Gastrointestinal Imaging • Original Research Sundarakumar et al. Image Registration for Subtractions in Liver DCE-MRI
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Gastrointestinal Imaging Original Research
Evaluation of Image Registration in Subtracted 3D Dynamic Contrast-Enhanced MRI of Treated Hepatocellular Carcinoma Dinesh K. Sundarakumar 1 Gregory J. Wilson Sherif F. Osman Sadaf F. Zaidi Jeffrey H. Maki Sundarakumar DK, Wilson GJ, Osman SF, Zaidi SF, Maki JH
Keywords: contrast agents, hepatocellular carcinoma, image registration, liver, MRI, registration DOI:10.2214/AJR.13.12417 Received December 11, 2013; accepted after revision June 27, 2014. Supported in part by Philips Healthcare. Based on a presentation at the Society of Computed Body Tomography and Magnetic Resonance 2013 annual meeting, Tucson, AZ. 1
All authors: Department of Radiology, University of Washington, 1959 NE Pacific St, Box 357115, Seattle, WA 98195-7115.. Address correspondence to D. K. Sundarakumar ([email protected]).
AJR 2015; 204:287–296 0361–803X/15/2042–287 © American Roentgen Ray Society
OBJECTIVE. The objective of our study was to quantify hepatic displacement between breath-holds in multiphasic contrast-enhanced MRI and assess the value of a 3D registration algorithm for displacement correction on subtracted images. MATERIALS AND METHODS. For this retrospective analysis, we evaluated MR images of 25 cirrhotic patients with treated hepatocellular carcinoma (HCC) and at least one coexisting small hepatic cyst that was hypointense on T1-weighted imaging. With the use of an automated 3D deformable registration algorithm, registered base and subtraction images were created using portal venous phase images as the baseline images. The relative displacement of the cysts over the dynamic phases was used to estimate hepatic displacement before and after registration. The width of the subtraction band artifact, HCC lesion conspicuity, and overall subtraction artifact level (i.e., image quality of the entire volume) of the subtraction images were evaluated before and after registration on a 5-point scale (1 = nondiagnostic, 5 = excellent image quality) by two blinded radiologists. Hepatic displacement and subtraction band artifact results were analyzed using the paired Student t test, and the results for HCC lesion conspicuity and image quality of the volume results were analyzed using the Wilcoxon signed rank test. Interobserver agreement was assessed using kappa statistics. RESULTS. The average total cyst displacement on unenhanced, arterial, and delayed phase images was significantly reduced by registration from 4.0, 3.2, and 4.6 mm, respectively, on preregistered images to 2.4, 1.6, and 1.3 mm on postregistered images (p < 0.01). The mean HCC lesion conspicuity grade improved from 3.4 before registration to 4.4 after registration (p < 0.01), and the mean grade for image quality of the volume improved from 3.3 before registration to 4.6 after registration (p < 0.01). The average width of the subtraction band artifact decreased from 5.3 mm before registration to 2.4 mm after registration, from 6.1 mm before registration to 2.6 mm after registration, and from 5.2 mm before registration to 2.8 mm after registration for the arterial, portal venous, and delayed phase subtractions, respectively (p < 0.01). CONCLUSION. Automated registration of the liver in multiphasic MRI examinations reduced interphasic hepatic displacement, improved the conspicuity of the treated HCC lesion, and improved the overall subtraction image quality.
S
ubtraction images are an integral part of routine multiphasic contrast-enhanced MRI for characterizing enhancement of lesions that are intrinsically hyperintense on T1weighted imaging, such as some hepatocellular carcinomas (HCCs), regenerative and dysplastic nodules, and many ablated HCCs [1]. In the latter case, T1 hyperintense foci often develop in the ablation zone after percutaneous ablation because of tissue necrosis, which complicates the detection of residual enhancement due to viable tumor. Accurate subtraction of the unenhanced phase removes preexisting T1 hyperintensity, thereby in-
creasing lesion contrast compared with the background liver and better facilitating the detection of subtle lesion enhancement. Thus, one crucial role for subtraction imaging is the detection and characterization of arterial enhancement in postablation HCCs, an imaging cornerstone for the diagnosis of residual or recurrent HCC [2]. Subtracted image datasets require precise 3D alignment to be free from registration artifacts. The liver undergoes considerable displacement and, to a lesser extent, elastic deformation due not only to respiratory but also to cardiac motion. Although breathing is cyclic, its critical endpoint—that is, the
AJR:204, February 2015 287
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al. Fig. 1—Schema of operations. Registration of other phases to portal venous phase results in postregistered series. Portal venous phase dataset is unchanged. Postregistered subtracted series generated by workstation uses postregistered unenhanced phase as mask image.
Preregistered series
Postregistered series
Unenhanced phase
Postregistered unenhanced phase
Registered to portal venous phase
Arterial phase
Postregistered arterial phase
Portal venous phase
Portal venous phase
Delayed phase
Postregistered delayed phase
Preregistered subtraction
Postregistered subtraction
Arterial phase
Portal venous phase
breath-hold—may vary from one acquisition to the next. This variation leads to variations in the position of the liver during examinations with repeated breath-holds and to subsequent misregistration on subtraction images [3]. As one would expect, the maximum respiratory variation in displacement has been shown to occur in the craniocaudal direction [4]. However, some anteroposterior motion and mediolateral motion also occur, and there is significant heterogeneity in the relative displacement of different liver segments during respiration [5]. This difference in relative displacements of different liver segments occurs because, in addition to inconsistencies in the depth of breath-holds,
Delayed phase
Arterial phase
cardiac pulsation (most pronounced in the left lobe) and elastic deformation add to the complexity of intersegmental hepatic movement [3]. Cardiac pulsation and elastic deformation result in additional in-plane misregistration for lesions near the dome and capsule of the liver where subtraction “band” artifacts are often seen [6]. Applying a simple purely z-axis (craniocaudal) motion correction by rigid transformation (slice shifting) may not offer a satisfactory solution to this perplexing problem of in-plane misregistration. The usefulness of subtraction images for characterizing small and subcapsular lesions has therefore been limited because of the resulting misregistration artifacts [1].
A
Portal venous phase
Delayed phase
Registration of hepatic motion is of clinical interest in multiphasic breath-hold examinations to reduce subtraction artifacts and improve lesion characterization [5]. Various image registration techniques have been described for multiphasic examinations of the liver [7]. The purpose of this study was to quantify the displacement of the liver between the different dynamic phases in a multiphasic examination, measure the improvement in displacement after registration using anatomic landmarks, and quantify the improvements in image quality and reader confidence for evaluating treated HCC. Intrahepatic cysts were used as anatomic landmarks, and registration was performed by a fully automated
B
Fig. 2—Measurement of subtraction band artifacts at right hepatic vein–inferior vena cava (IVC) junction in 43-year-old woman with treated hepatocellular carcinoma. A and B, Preregistered (A) and postregistered (B) arterial phase images show decrease in maximum thickness of band artifact for all four quadrants. Vertical dashed line passes through mid IVC, and horizontal dashed line passes 2 cm anterior to IVC. Greatest band thickness perpendicular to liver surface (solid lines) is measured in each quadrant.
288
AJR:204, February 2015
The local institutional review board approved this HIPAA-compliant retrospective study and waived the requirement for informed consent. The data from 25 multiphasic liver MRI studies, all performed to evaluate treated HCC in cirrhotic patients (20 men and five women; age range, 35–72 years) scanned between January 2010 and December 2011, were analyzed. These patients were chosen for this study because, in addition to treated HCC, they also had small easily seen intrahepatic cysts; a total of 32 T1 hypointense cysts conspicuous in all dynamic phases were used as intrahepatic markers to measure displacement without registration and with registration. Each patient’s treated HCC lesion (n = 25) was evaluated for lesion conspicuity on nonregistered and registered subtraction images.
MR Image Acquisition Multiphasic contrast-enhanced MRI was performed on a 1.5-T scanner (Achieva, Philips Healthcare) using a 16-channel torso phased-array coil. An axial 3D two-echo T1 fast-field echo modified Dixon (Philips Healthcare) sequence was used with the following parameters: TR/TE1, TE2, 6.0/1.8, 4.0; flip angle, 10°; FOV, 38–40 cm; matrix size, 204–320 × 192–208; and scanning time per dynamic phase, 15–18 seconds [8]. Imaging was performed before and after IV injection of 0.1 mL/ kg of gadoteridol (ProHance, Bracco Diagnostics) using an MRI-compatible power injector (Spectris, Medrad) in end-expiratory breath-hold (one dynamic phase per breath-hold). Contrast-enhanced scans were manually triggered after fluoroscopic bolus arrival detection at the aortic bifurcation with approximate acquisition start times of 30 seconds (arterial phase), 70 seconds (portal venous phase), and 5 minutes (delayed phase) after injection.
Image Registration and Subtraction The breath-hold 3D modified Dixon images were transferred to a U.S. Food and Drug Administration–approved dedicated workstation (CADstream liver workstation, version 5.3.1.218, Merge Healthcare). Registered base and subtraction images were created using the workstation’s fully automated 3D deformable registration algorithm (Fig. 1). This propriety software, according to its default settings, registers unenhanced, arterial, and equilibrium phase data to the portal venous phase. The registration algorithm incorporates a fully automated combination of rigid and isotropic elastic nonrigid transformation using featurebased detection and matching methods that are in-
5.0
Reader 2
Reader 3
4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0
Arterial Phase
Portal Venous Phase
Delayed Phase
Arterial Phase
Preregistered Series
Portal Venous Phase
Delayed Phase
Postregistered Series
A 6.0 Reader 2
Reader 3
5.0 4.0 3.0 2.0 1.0 0
Arterial Phase
Portal Venous Phase
Delayed Phase
Arterial Phase
Preregistered Series
Portal Venous Phase
Delayed Phase
Postregistered Series
B 7 6
Fig. 3—Analyses of image quality of pre- and postregistered subtraction images. A, Bar graph shows increase in subjective image quality of conspicuity of treated hepatocellular carcinoma after registration (1 = nondiagnostic, 5 = excellent image quality). HCC = hepatocellular carcinoma. B, Bar graph shows increase in subjective image quality of entire subtraction volume after registration (1 = nondiagnostic, 5 = excellent image quality). C, Bar graph shows decrease in mean width of subtraction band artifact after registration.
dependent of the image intensity and therefore not influenced by implicit signal intensity data, which is the actual variable we are most interested in measuring in hepatic MRI [9]. The technical details of this registration algorithm can be found in the following references: [7, 10].
Mean Width of Subtraction Band Artifact (mm)
Materials and Methods Subjects
Subjective Grade for Image Quality of Entire Subtraction Volume
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
nonrigid transformation algorithm on a commercially available liver workstation.
Subjective Grade for Conspicuity of Treated HCC
Image Registration for Subtractions in Liver DCE-MRI
5 4 3 2 1 0
Arterial Phase
Portal Venous Phase
Preregistered series
Delayed Phase
Postregistered series
C Subtracted datasets were created by subtracting the unenhanced (mask) images from each of the contrast-enhanced phases (arterial, portal venous, and delayed). Subtraction of the nonregistered datasets resulted in the data henceforth referred to as “preregistered subtraction” data. Subtraction of
AJR:204, February 2015 289
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al.
A
B
C
D
E
F
Fig. 4—45-year-old man with treated hepatocellular carcinoma HCC in segment VIII. Poor image quality affects image registration and lesion depiction. A and B, Preregistered axial T1-weighted arterial (A) and delayed (B) images show blurring of liver margins and lesion margins (arrow). C and D, Postregistered axial arterial (C) and delayed (D) images show accentuation of blurring of lesion margins (arrow). Readers rated (side-by-side image degradation) preregistered source images as better than postregistered source images for depiction of HCC. Postregistration source image degradation occurred in five of 25 patients. E and F, Preregistered (E) and postregistered (F) subtracted arterial phase images. Distortion and blurring (arrow) are more pronounced in postregistered subtracted arterial phase image than in preregistered subtracted arterial phase image. Reduction of subtracted image quality (i.e., of lesion conspicuity and image quality of volume) after registration occurred for this patient only.
the registered datasets—henceforth referred to as “postregistered subtraction”—was performed after source image registration.
ments in craniocaudal, anteroposterior, and mediolateral directions were also evaluated for each image set. In livers with multiple cysts, the average cyst displacement was used for the analysis.
Displacement Measurements Intrahepatic subcentimeter T1 hypointense cysts detectable on all phases were chosen as discrete intrahepatic landmarks and were used to calculate hepatic displacement during multiphasic scanning. The center voxel location of each T1 hypointense focus was determined with respect to x-, y-, and z-axis coordinates (mediolateral, anteroposterior, and craniocaudal, respectively) on each pre- and postregistered dynamic image set by one blinded radiologist who had 5 years’ experience at the time of the study (reader 1). Euclidean displacement (E) in the unenhanced (U), arterial (A), and delayed (D) phases was calculated using the following formula: EU,A,D = {(XP – XU,A,D)2 + (YP – YU,A,D)2 + (ZP – ZU,A,D)2} where X, Y, and Z are the x-, y-, and z-axis coordinates in 3D; and the subscripts U, A, P, and D denote unenhanced, arterial, portal venous, and delayed phases, respectively. Single-axis displace-
290
Image Evaluation Randomized subjective analysis—Patient-identifying information was removed from the image sets, and the image sets were sent to the hospital PACS for quality analysis. These datasets were presented in a random manner to two blinded radiologists (reader 2 with 5 years’ experience and reader 3 with 4 years’ experience) who were unaware of the registration status. They independently analyzed the 25 treated HCCs on both pre- and postregistered subtraction images and assessed these images in two ways. First, the readers evaluated the subtracted images for diagnostic confidence in treated HCC depiction and enhancement pattern (HCC lesion conspicuity); henceforth, we refer to this parameter as “lesion conspicuity.” The readers assessed lesion conspicuity using the following 5-point scale: 1, nondiagnostic, lesion not visualized; 2, lesion barely visualized with enhancement fea-
tures not assessed because of severe artifacts; 3, lesion visualized with uncertain enhancement features; 4, lesion visualized with sufficiently discernible enhancement pattern and minimal artifacts or slight blurring of the edges; and 5, lesion visualized with well-characterized enhancement pattern and sharp margins. Grades 4 and 5 were considered acceptable image quality. Second, the readers characterized their overall level of confidence in subtraction quality by taking into account the entire subtraction volume as a whole, which we refer to henceforth as “image quality of the volume.” The image quality of the volume was graded using the following 5-point scale: 1, overall nondiagnostic image quality; 2, severe subtraction artifacts (maximum thickness of subtraction bands of > 5 mm or more than two thirds of either the liver margins or vascular markings are ill defined); 3, moderate subtraction artifacts (maximum thickness of subtraction bands of 2–5 mm or between one third and two thirds of either the liver margins or vascular markings are ill defined); 4, good overall image quality with minimal artifacts (maximum thickness of subtraction bands of < 2 mm or less than one third of either
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI the liver margins or vascular markings are ill defined); and 5, perfect overall subtraction quality without any artifacts. Grades 4 and 5 were considered satisfactory image quality. Side-by-side analysis—The maximum perpendicular thickness of the subtraction band artifact was measured by reader 1 in each of four quadrants on the liver surface at the level of the junction of the hepatic veins with the inferior vena cava in both the pre- and postregistered subtraction images (Fig. 2). The cysts, whose true enhancement nature was known (i.e., none), were chosen as targets to assess the burden of potential false-positive enhancement on the characterization of focal liver lesions. Focal false-positive enhancement in the subtracted images due to misregistration of the T1 hypointense cysts was measured by reader 1 in a binary fashion (yes or no). The preregistered source images were used as a reference standard to confirm or refute any findings of enhancement in these foci. Finally, to assess for potential source image degradation due to the image registration process, two radiologists (readers 2 and 3) reviewed the pre- and postregistered source images (blinded as to which) side-by-side to compare HCC lesion conspicuity and margin definition. The following question about the preregistered (A) and postregistered (B) series was asked: On which image set is the HCC lesion best seen (image series A or image series B), or were both image series equally good?”
Statistical Analysis The displacement of T1 hypointense foci and the width of the subtraction band artifact were analyzed using the two-way paired Student t test. The overall diagnostic confidence and image artifacts (lesion conspicuity and image quality of the volume) were compared between pre- and postregistered subtraction images using the Wilcoxon signed rank test. Evaluation of reader preference for preversus postregistered source images in terms of image degradation was done by Monte Carlo simulation of multinomial data. Interobserver agreement was assessed by Cohen’s kappa statistics.
Results Lesions in Study Population Two patients had three liver cysts used to calculate displacement, three patients had two cysts, and all others had a solitary liver cyst. Twenty-five treated HCCs (average size, 2.8 cm; range, 1.5–8.9 cm) were analyzed (one HCC per patient). These lesions were located in segments II (n = 2), III (n = 2), IVa (n = 4), V (n = 1), VI (n = 4), VII (n = 6), and VIII (n = 6). Thirteen treated HCCs were intrinsically hyperintense on T1weighted imaging, and 12 were hypointense on T1-weighted imaging. Seventeen of 25 HCCs (68%) were subcapsular in location, and eight (32%) lesions were central in location. Fifteen HCCs (60%) were located in
TABLE 1: Measured Displacements of Anatomic Landmarks Mean (SD) [Range] Displacement of Anatomic Landmark (mm) Displacement Measurement Method and Imaging Phase
Preregistered Subtraction Images
Postregistered Subtraction Images
p
Euclidian displacement Unenhanced
4.0 (2.6) [1.1–11.9]
2.4 (1.4) [0.1–5.3]
0.006
Arterial
3.2 (3.5) [0.0–18.0]
1.6 (1.4) [0.0–6.3]
0.004
Delayed
4.6 (3.7) [1.1–14.9]
1.3 (1.0) [0.0–5.1]
< 0.001
Unenhanced
2.5 (2.2) [0.0–8.8]
1.1 (1.6) [0.0–4.0]
< 0.001
Arterial
2.2 (2.7) [0.0–14.0]
0.8 (1.2) [0.0–4.0]
0.002
Delayed
2.9 (3.0) [0.0–12.0]
0.5 (1.0) [0.0–4.0]
< 0.001
Unenhanced
1.3 (1.4) [0.0–0.6]
1.0 (0.8) [0.0–4.0]
0.16
Arterial
0.9 (0.8) [0.0–3.3]
0.5 (0.4) [0.0–2.1]
0.02
Delayed
1.6 (1.2) [0.0–4.3]
0.6 (0.6) [0.0–2.7]
< 0.001
Unenhanced
1.5 (1.8) [0.0–10.1]
1.0 (0.9) [0.0–3.3]
0.11
Arterial
1.5 (2.1) [0.0–11.2]
0.8 (0.9) [0.0–4.3]
0.02
Delayed
1.6 (2.7) [0.0–13.3]
0.6 (0.6) [0.0–1.9]
0.04
Craniocaudal displacement
Mediolateral displacement
Anteroposterior displacement
the same segment as the cyst that was used as the marker for motion evaluation. The average distance of the cyst from the HCC was 3.9 cm (range, 1.7–10.1 cm). Residual disease was detected in two of these treated HCCs. Motion Assessment The mean displacements of the anatomic landmarks (i.e., the cysts) are listed in Table 1. The average Euclidian displacement was significantly lower after registration for all three phases (pre- vs postregistered measurements: unenhanced, 4.0 vs 2.4 mm, respectively; arterial, 3.2 vs 1.6 mm; delayed, 4.6 vs 1.3 mm). Craniocaudal displacement was greatest: The average craniocaudal displacement over the three phases was 2.6 mm preregistered and 0.8 mm postregistered. The improvement in craniocaudal displacement was significant for each of the three phases. The average displacement of all three phases in the anteroposterior and mediolateral directions were 1.5 and 1.3 mm preregistered, decreasing to 0.8 and 0.7 mm postregistered. The decrease in anteroposterior and mediolateral displacement after registration was significant for the arterial and delayed phases. Randomized Subjective Analysis Evaluation of the subtracted image sets showed there was a statistically significant increase in lesion conspicuity after image registration (average pre- vs postregistered grade, 3.4 and 4.4 respectively; p < 0.01) (Table 2 and Fig. 3A). The average percentage of satisfactory subtraction image sets (lesion conspicuity grades 4 and 5) increased with registration—from 38% of preregistered subtracted image sets to 90% after registration (Table 2 and Fig. 3A). The global subtraction artifact level (image quality of the volume) also significantly improved in postregistered subtractions (3.3 vs 4.6 for pre- and postregistered subtractions, respectively; p < 0.01). Furthermore, the average percentage of satisfactory subtraction image sets (image quality of volume grades 4 and 5) increased with registration—from 34% of preregistered subtracted image sets to 95% after registration (Table 2 and Fig. 3B). As can be seen from Table 2, the kappa values for the two readers range from 0.6 to 0.8, reflecting good to excellent agreement between the two readers. Side-by-Side Analysis The average width of the subtraction band artifact decreased with registration from 5.3 to
AJR:204, February 2015 291
Sundarakumar et al.
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
TABLE 2: Evaluation of Subtracted Images
Parameter Assessed and Imaging Phase
Preregistered Subtraction Images
κ
Postregistered Subtraction Images
κ
HCC lesion conspicuity, mean grade [no. of cases with grade of 1, 2, 3, 4, 5, respectively] Arterial phase
0.8a
0.7
Reader 2
3.4 [0, 0, 15, 9, 1]
4.4 [0, 1, 1, 11, 12]
Reader 3
3.4 [0, 0, 14, 10, 1]
4.4 [0, 1, 0, 12, 12]
Portal venous phase
0.6a
0.7
Reader 2
3.5 [0, 0, 14, 9, 2]
4.5 [0, 1, 1, 7, 16]
Reader 3
3.3 [0, 0, 16, 8, 1]
4.4 [0, 1, 2, 7, 15]
Delayed phase
0.6a
0.6
Reader 2
3.4 [0, 0, 15, 9, 1]
4.5 [0, 1, 3, 3, 18]
Reader 3
3.1 [0, 0, 19, 5, 1]
4.5 [0, 1, 2, 4, 18]
3.4
4.4
Average
a
Image quality of volume, mean grade [no. of cases with grade of 1, 2, 3, 4, 5, respectively] Arterial phase
0.6a
0.7
Reader 2
3.1 [0, 3, 16, 5, 1]
Reader 3
3.1 [0, 3, 14, 7, 1]
Portal venous phase
4.4 [0, 1, 0, 10, 14] 4.4 [0, 1, 0, 12, 12] 0.6a
0.7
Reader 2
3.5 [0, 0, 13, 10, 2]
4.8 [0, 1, 0, 2, 22]
Reader 3
3.4 [0, 0, 15, 10, 0]
4.7 [0, 1, 0, 4, 20]
Delayed phase
0.7a
0.8
Reader 2
3.4 [0, 2, 15, 5, 3]
4.7 [0, 1, 0, 5, 19]
Reader 3
3.3 [0, 2, 16, 5, 2]
4.7 [0, 1, 1, 3, 20]
3.3
4.6
Arterial phase
5.3 (3.7)
2.4 (2.1)
Portal venous phase
6.1 (6.6)
2.6 (1.7)
Delayed phase
5.2 (4.9)
2.8 (3.0)
Average
5.5 (4.6)
2.6 (2.2)
10/32
0/32
Average
a
Subtraction band width (mm), mean (SD)
False-positive enhancement, no. of cysts/total no. of cysts
a
ap < 0.01.
2.4 mm in the arterial phase subtractions, from 6.1 to 2.6 mm in the portal venous phase subtractions, and from 5.2 to 2.8 mm in the delayed phase subtractions (p < 0.01) (Table 2 and Fig. 3C). However, in one case the postregistered subtraction dataset was worse than the preregistered subtraction dataset (Fig. 4). In the evaluation for the potential effects of image degradation on lesion conspicuity between pre- and postregistered source images, reader 2 rated the preregistered source images as better in five cases, worse in two cases, and equal in 18 cases. Reader 3 had almost identical results, rating the preregistered source images better in
292
four cases, worse in three cases, and equal in 18 cases (κ = 0.3). This trend toward source image degradation after registration did not reach statistical significance (p > 0.07). Ten of 32 cysts (31%) showed false-positive enhancement due to misregistration on the preregistered subtracted images. No false-positive enhancement due to cyst misregistration was seen in the postregistered subtraction datasets (Fig. 5). Discussion The results of this study revealed that considerable hepatic displacement occurred be-
tween breath-holds in all patients and, as expected, was greatest in the craniocaudal direction. These results almost certainly reflect the predominately craniocaudal inconsistencies in diaphragmatic (and therefore liver) positioning during repeated breathholds. Investigators noted in a previous study that quiet breathing results in craniocaudal displacement ranging from 10 to 26 mm [11]. Our measurements of displacement were considerably less, averaging 2.6 mm (maximum, 14.0 mm) and provides good estimates of variations of hepatic positions between multiple expiratory breath-holds in clinical abdominal MRI. The mean displacement was greatest on the delayed phase. This greater displacement likely results from the lengthy delay (≈ 5 minutes) between the earlier phases (relatively rapid back-to-back unenhanced, arterial, portal venous phases) and the delayed phase, at which time patients may not remember the “feel” of that same breath-hold or may have moved. After registration, the mean displacement decreased in all three directions, with average Euclidian displacement decreasing from 3.9 to 1.8 mm; average craniocaudal displacement, from 2.6 to 0.8 mm; average mediolateral displacement, from 1.3 to 0.7 mm; and average anteroposterior displacement, from 1.5 to 0.8 mm. The reductions in Euclidian and craniocaudal displacements were statistically significant in all three dynamic phases. In one extreme instance, we noted a maximum craniocaudal displacement of 14.0 mm (Euclidian displacement, 18.0 mm) in the arterial phase that after registration corrected to 4 mm (Euclidian displacement, 6.25 mm). Larger displacements, particularly those confined to the craniocaudal axis could, in principle, be registered by a purely rigid transformation. Unlike craniocaudal translation, however, the mediolateral and anteroposterior displacements (which are similar in magnitude) more likely reflect elastic deformation and therefore depend predominantly on nonrigid transformation for motion correction. Using the rigid and isotropic elastic nonrigid algorithm under investigation, we found that both anteroposterior and mediolateral displacements were also reduced by registration, achieving significance for anteroposterior and mediolateral displacements in the arterial and delayed phases. Registration of subtraction images led to considerable improvement in subjective HCC lesion conspicuity. There was also a significant decrease in subtraction band ar-
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI
A
B
C
D
Fig. 5—55-year-old woman with treated hepatocellular carcinoma (HCC) in segment IVa. False-positive enhancement of cyst was seen on preregistered subtraction image. A, Axial T2-weighted image shows cyst in segment VII (arrowhead); this cyst was used to measure hepatic motion. B, Axial unenhanced T1-weighted image shows heterogeneous hyperintensity in ablated cavity (arrow) and same hypointense cyst in segment VII (arrowhead). C, Axial arterial phase preregistered subtracted image (overall image quality of entire subtraction volume, grade 3) shows nodular and heterogeneous enhancement in periphery (dashed arrow) and uncertain enhancement within ablation cavity (solid arrow) (lesion conspicuity, grade 3). False-positive enhancement (arrowhead) due to misregistration is seen in segment VII cyst. D, Postregistered subtraction image in arterial phase at same level as C (overall image quality of entire subtraction volume, grade 4) shows linear enhancement limited to ablation margins and no enhancement within ablated cavity (solid arrow) (lesion conspicuity, grade 5). There is no misregistration-related enhancement (i.e., falsepositive enhancement) in segment VII cyst (arrowhead). Significant reduction in band artifact is also noted, particularly along inferior border of liver (dashed arrow).
tifacts along the liver margin after registration, allowing improved assessment of the subcapsular regions of liver, especially in the dome and inferior margins. This decrease in subtraction band artifacts is particularly advantageous for characterizing subcapsular lesions (Fig. 6). Similarly, the overall subtraction artifacts (image quality of the volume) significantly decreased within the liver because of the associated decreased misregistration of intrahepatic structures such as vessels. This decrease in artifacts may in part explain the favorable trend toward improvement in lesion conspicuity; the percentage of subtraction images rated satisfactory before registration (38%; average total displacement, 3.9 mm) more than doubled after registration to 90% (average total displace-
ment, 1.8 mm) (Table 2). There was only one case in which the subtraction image quality (both lesion conspicuity and image quality of the volume) was worse after registration (Fig. 4). The decrease in subtraction image quality in the latter case was likely a result of patient motion during acquisition, which caused blurring of the hepatic margins and internal features and led to increased artifacts in the preregistered images. In this case, registration based on compromised edge definition caused further deterioration of the image quality due to the smoothing filters used in the algorithm. In all other cases, registration was successful in mitigating the misregistration in the subtracted images. Yu et al. [1] observed that for preregistered subtracted images, the enhancement
features of lesions larger than 2 cm could always be reliably characterized, whereas the enhancement features could be reliably characterized in only 51% of lesions smaller than 2 cm. With the improvement in imaging technology, the minimum size of a nodule for which the enhancement criteria for diagnosing HCC on CT or MRI are applicable has decreased from 2 to 1 cm [12, 13]. Therefore, as characterizable lesions become smaller, we anticipate that the deleterious effects of misregistration artifacts in subtraction imaging may become more pronounced. In this study, 10 HCC lesions were equal to or smaller than 2 cm. Both readers characterized images from postregistered subtracted datasets of eight of these lesions (80%) as being of acceptable diagnostic quality—that is, without significant
AJR:204, February 2015 293
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al.
A
B
C
D
Fig. 6—60-year-old woman with treated hepatocellular carcinoma (HCC). Presence of band artifact affects characterization of subcapsular lesions. A and B, Preregistered subtracted arterial (A) and portal venous (B) images show misregistration of perinephric fat on segment VI (overall image quality of entire subtraction volume, grade 3) causing false-positive (in venous phase) enhancement in treated HCC (solid arrow, B) and obscuring small cyst (dashed arrow, B) in same segment (lesion conspicuity, grade 3). C and D, No band artifacts are noted in postregistered subtracted arterial (C) and portal venous (D) images (overall image quality of entire subtraction volume, grade 5). Thin linear enhancement (solid arrow) is noted around treated HCC (lesion conspicuity, grade 5) in arterial and portal venous phases and no enhancement is noted in cyst (dashed arrow, D). Note emergence of coherent artifacts in postregistered images because feature-based image registration may also boost depiction of artifacts (e.g., motion) present in source images.
artifacts (lesion conspicuity grade 4 or 5). The conspicuity of all of these lesions was assessed as grade 3 on preregistered subtraction series. Thus, in eight of 10 lesions that were equal to or smaller 2 cm, the image quality was improved from clinically nonsatisfactory (lesion conspicuity grade 3 or lower) to clinically satisfactory (lesion conspicuity grade 4 or 5). Although the focus of this study was not the detection and full characterization of HCC lesions, this improved image quality potentially opens new frontiers for more accurate characterization of small (i.e., 1–2 cm) lesions; this need to characterize small lesions arises in part from the use
294
of the management-based Liver Imaging Reporting and Data System (LI-RADS) classification [14]. In the previously mentioned study, Yu et al. [1] also found that the quality of the subtraction images was significantly better for images of central lesions than for images of subcapsular lesions, the latter of which they observed were often obscured by band artifacts. In our study, although a separate analysis based on the location and size of the lesion was not performed, 17 of 25 (68%) HCC lesions were subcapsular, and most of the subcapsular HCC lesions were well characterized. The observations of both readers
showed that the lesion conspicuity grade after registration improved from 3 to 4 or 5 for 14 of the 17 subcapsular lesions, improved from 4 to 5 in two lesions, and remained grade 5 in one lesion because of the significant reduction in subtraction artifacts in the subcapsular region of the liver (Fig. 6). This decrease in artifacts may in part explain the overall trend toward significant improvement in lesion depiction (lesion conspicuity) on postregistered subtraction images, with the limitations mentioned earlier at least partly overcome by registration. Moreover, any false-positive enhancement of the cysts on the subtraction images resolved after regis-
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI tration, further suggesting the potential for improvement in the evaluation of small hepatic lesions [3]. When compared with landmark- or surface-based registration, the registration method used in this study is robust for a highly mobile and deformable organ such as the liver. The feature-based registration methods used are based on similarities between corresponding points in the images. This technique is not affected by signal intensity differences, which occur during various phases of multiphasic scanning because of contrast enhancement. Some earlier enhancement-driven methods have used iterative registration generated by fitting a predetermined model of contrast pharmacokinetics to the datasets. This approach requires complex models and is biased toward the model itself and therefore may not truly reflect the enhancement pattern in all cases [7]. However, 3D deformable registration, as evaluated in this study, can in principle alter or degrade the appearance of lesions. Although the data show that subtraction image quality is improved by registration, it is important to also consider whether lesions may be degraded by the deformation that may occur during registration (image degradation). In this small evaluation of only treated HCCs, reader 2 rated five preregistered source image datasets as superior, two as inferior, and 18 as equal to the postregistered image datasets. Similarly, reader 3 rated four preregistered source image datasets as superior, three as inferior, and 18 as equal to the postregistered image datasets. These observations had poor interobserver agreement and were not statistically significant. Some reduction in image quality was likely caused by the inherent effects of smoothing, which occur during registration and change the appearance of the lesions. Although the effects of smoothing were too subtle to detect in 18 of 25 patients, it would be pertinent to ultimately assess the true dimensions, edge sharpness, and contrast of the lesions in the preregistered images alongside the postregistered images to investigate any potential discrepancy caused by image distortion after registration. This assessment was beyond the scope of this study, so this issue remains an important consideration if lesion diagnosis were to be made exclusively from the registered source images and bears further investigation. We should also note that although registration corrects the position variation occurring at the time interval between mul-
tiphasic scans, the registration is particularly prone to the deleterious effects of motion occurring during the scan itself. This motion is likely caused by motion artifacts in the source images that impair the feature-based matching process occurring in registration. In this study, most source images were devoid of significant artifacts (high overall image quality); thus, the subtraction algorithm had a higher likelihood of success. Although not evaluated in this study, a dedicated workstation approach to liver evaluation offers simultaneous display of anatomically synchronized multiphasic datasets despite wide variations in the actual positioning between them. For example, the craniocaudal displacement in preregistered images averaged 2.6 mm with a maximum of 14.0 mm. This amount of displacement means that 2-mm-thick axial images depicting a lesion could be potentially seven slices apart on different phases. After registration, with an average craniocaudal displacement of 0.8 mm and a maximum of 4.4 mm, this disparity would be at a maximum of two slices. This smaller disparity facilitates automatic linking and scrolling ability and allows a simplified analysis of enhancement kinetics (e.g., an ROI cursor placed in one postregistered phase is better aligned with the other phases). Dedicated liver workstation interfaces can also include functions such as apparent diffusion coefficient mapping, contrast kinetics, and volumetric analysis in addition to signal intensity characterization. The potential scope of improvement in workflow underscores its importance in a busy hepatic imaging clinic. Although the time required to transfer the study to a workstation, execute image registration, and transfer the postregistered images back to the PACS and its influence on workflow were not systematically analyzed (a function of the retrospective nature of this project), our datasets were exported from a PACS and processed on the workstation in approximately 5–7 minutes per case. Note that this process can occur in the background while the radiologist attends to other matters, and the images can then be analyzed at any time. This study has several limitations. First, the sample size was relatively small for achieving a large-scale validation of the robustness of registration and for justifying the added resources (e.g., time, equipment) required. However, this study was designed to measure the performance of automated registration in the delineation of treated HCC
and determine whether there is potential benefit to registered subtraction images, both of which were done. Second, the displacement of a small number of discrete focal points (cysts) within the liver was evaluated as a simplified estimate of linear translation. The use of this strategy for our study means that rotational and nonrigid deformations were not directly measured; therefore, these results do not give us insights into the degree of elastic deformation that is occurring. Third, because of significant intersegmental heterogeneity in hepatic displacement during respiration, the cysts used as internal reference points for displacement may not truly reflect the displacement of the HCC if they were not situated in near proximity. Fourth, there was no objective means by which to measure the effects of registration-associated smoothing and distortion on image quality. Instead, we subjectively assessed for any HCC degradation by blinded side-by-side comparison of nonsubtracted images, and most cases showed no artifacts to compromise HCC lesion evaluation. As we previously stated, however, this strategy needs to be investigated further before diagnosis is performed exclusively from registered datasets. Finally, the time required to transfer the study to the workstation, execute image registration, and transfer the postregistered images back to the PACS and the influence of this time investment on workflow were not systematically analyzed (a function of the retrospective nature of this project). However, depending on server and workstation speed, our datasets were exported and processed on the workstation with an acceptable time penalty (≈ 5–7 minutes per case). This process can occur in the background while the radiologist attends to other matters, and the images can then be analyzed at any time. In conclusion, this study quantified liver displacement between sequential expiratory breath-holds in multiphasic MRI and showed significant reductions in these displacements after automated 3D image registration. A commercially available, fully automated nonrigid registration algorithm significantly reduced subtraction artifacts and improved diagnostic confidence in the assessment of treated HCC by correcting image misregistration due to inconsistent liver positioning during sequential breath-holds. This capability has potential implications for the detection and characterization of small (≤ 2 cm) and subcapsular lesions, which are traditionally difficult to evaluate with subtraction because of artifacts. These
AJR:204, February 2015 295
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al. preliminary results suggest registration may be helpful in routine liver MRI, particularly for follow-up of HCCs after ablation. Further work is required to evaluate the full impact of registration on lesion characterization. References 1. Yu JS, Kim YH, Rofsky NM. Dynamic subtraction magnetic resonance imaging of cirrhotic liver: assessment of high signal intensity lesions on nonenhanced T1-weighted images. J Comput Assist Tomogr 2005; 29:51–58 2. Kuehl H, Rosenbaum-Krumme S, Veit-Haibach P, et al. Impact of whole-body imaging on treatment decision to radio-frequency ablation in patients with malignant liver tumors: comparison of [18F] fluorodeoxyglucose-PET/computed tomography, PET and computed tomography. Nucl Med Commun 2008; 29:599–606 3. Yu JS, Rofsky NM. Dynamic subtraction MR imaging of the liver: advantages and pitfalls. AJR 2003; 180:1351–1357 4. Clifford MA, Banovac F, Levy E, Cleary K. As-
sessment of hepatic motion secondary to respiration for computer assisted interventions. Comput Aided Surg 2002; 7:291–299 5. Srimathveeravalli G, Leger J, Ezell P, Maybody M, Gutta N, Solomon SB. A study of porcine liver motion during respiration for improving targeting in image-guided needle placements. Int J Comput Assist Radiol Surg 2013; 8:15–27 6. Tatli S, Acar M, Tuncali K, Sadow CA, Morrison PR, Silverman SG. MRI assessment of percutaneous ablation of liver tumors: Value of subtraction images. J Magn Reson Imaging 2013; 37:407–413 7. Melbourne A, Atkinson D, White MJ, Collins D, Leach M, Hawkes D. Registration of dynamic contrast-enhanced MRI using a progressive principal component registration (PPCR). Phys Med Biol 2007; 52:5147–5156 8. Eggers H, Brendel B, Duijndam A, Herigault G. Dual-echo Dixon imaging with flexible choice of echo times. Magn Reson Med 2011; 65:96–107 9. Melbourne A, Atkinson D, Hawkes D. Influence of organ motion and contrast enhancement on image registration. Med Image Comput Comput As-
sist Interv 2008; 11(pt 2):948–955 10. Rajaraman S, Rodriguez JJ, Graff C, et al. Automated registration of sequential breath-hold dynamic contrast-enhanced MR images: a comparison of three techniques. Magn Reson Imaging 2011; 29:668–682 11. Suramo I, Päivänsalo M, Myllylä V. Cranio-caudal movements of the liver, pancreas and kidneys in respiration. Acta Radiol Diagn (Stockh) 1984; 25:129–131 12. Forner A, Vilana R, Ayuso C, et al. Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: Prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma. Hepatology 2008; 47:97–104 13. Song do S, Bae SH. Changes of guidelines diagnosing hepatocellular carcinoma during the last tenyear period. Clin Mol Hepatol 2012; 18:258–267 14. Purysko AS, Remer EM, Coppa CP, Leão Filho HM, Thupili CR, Veniero JC. LI-RADS: a casebased review of the new categorization of liver findings in patients with end-stage liver disease. RadioGraphics 2012; 32:1977–1995
F O R YO U R I N F O R M AT I O N
Unique customized medical search engine service from ARRS! ARRS GoldMiner® is a keyword- and concept-driven search engine that provides instant access to radiologic images published in peer-reviewed journals. ARRS members earn 0.5 CME credits for each self-directed search. For more information, visit goldminer.arrs.org.
296
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Gastrointestinal Imaging Original Research
Evaluation of Image Registration in Subtracted 3D Dynamic Contrast-Enhanced MRI of Treated Hepatocellular Carcinoma Dinesh K. Sundarakumar 1 Gregory J. Wilson Sherif F. Osman Sadaf F. Zaidi Jeffrey H. Maki Sundarakumar DK, Wilson GJ, Osman SF, Zaidi SF, Maki JH
Keywords: contrast agents, hepatocellular carcinoma, image registration, liver, MRI, registration DOI:10.2214/AJR.13.12417 Received December 11, 2013; accepted after revision June 27, 2014. Supported in part by Philips Healthcare. Based on a presentation at the Society of Computed Body Tomography and Magnetic Resonance 2013 annual meeting, Tucson, AZ. 1
All authors: Department of Radiology, University of Washington, 1959 NE Pacific St, Box 357115, Seattle, WA 98195-7115.. Address correspondence to D. K. Sundarakumar ([email protected]).
AJR 2015; 204:287–296 0361–803X/15/2042–287 © American Roentgen Ray Society
OBJECTIVE. The objective of our study was to quantify hepatic displacement between breath-holds in multiphasic contrast-enhanced MRI and assess the value of a 3D registration algorithm for displacement correction on subtracted images. MATERIALS AND METHODS. For this retrospective analysis, we evaluated MR images of 25 cirrhotic patients with treated hepatocellular carcinoma (HCC) and at least one coexisting small hepatic cyst that was hypointense on T1-weighted imaging. With the use of an automated 3D deformable registration algorithm, registered base and subtraction images were created using portal venous phase images as the baseline images. The relative displacement of the cysts over the dynamic phases was used to estimate hepatic displacement before and after registration. The width of the subtraction band artifact, HCC lesion conspicuity, and overall subtraction artifact level (i.e., image quality of the entire volume) of the subtraction images were evaluated before and after registration on a 5-point scale (1 = nondiagnostic, 5 = excellent image quality) by two blinded radiologists. Hepatic displacement and subtraction band artifact results were analyzed using the paired Student t test, and the results for HCC lesion conspicuity and image quality of the volume results were analyzed using the Wilcoxon signed rank test. Interobserver agreement was assessed using kappa statistics. RESULTS. The average total cyst displacement on unenhanced, arterial, and delayed phase images was significantly reduced by registration from 4.0, 3.2, and 4.6 mm, respectively, on preregistered images to 2.4, 1.6, and 1.3 mm on postregistered images (p < 0.01). The mean HCC lesion conspicuity grade improved from 3.4 before registration to 4.4 after registration (p < 0.01), and the mean grade for image quality of the volume improved from 3.3 before registration to 4.6 after registration (p < 0.01). The average width of the subtraction band artifact decreased from 5.3 mm before registration to 2.4 mm after registration, from 6.1 mm before registration to 2.6 mm after registration, and from 5.2 mm before registration to 2.8 mm after registration for the arterial, portal venous, and delayed phase subtractions, respectively (p < 0.01). CONCLUSION. Automated registration of the liver in multiphasic MRI examinations reduced interphasic hepatic displacement, improved the conspicuity of the treated HCC lesion, and improved the overall subtraction image quality.
S
ubtraction images are an integral part of routine multiphasic contrast-enhanced MRI for characterizing enhancement of lesions that are intrinsically hyperintense on T1weighted imaging, such as some hepatocellular carcinomas (HCCs), regenerative and dysplastic nodules, and many ablated HCCs [1]. In the latter case, T1 hyperintense foci often develop in the ablation zone after percutaneous ablation because of tissue necrosis, which complicates the detection of residual enhancement due to viable tumor. Accurate subtraction of the unenhanced phase removes preexisting T1 hyperintensity, thereby in-
creasing lesion contrast compared with the background liver and better facilitating the detection of subtle lesion enhancement. Thus, one crucial role for subtraction imaging is the detection and characterization of arterial enhancement in postablation HCCs, an imaging cornerstone for the diagnosis of residual or recurrent HCC [2]. Subtracted image datasets require precise 3D alignment to be free from registration artifacts. The liver undergoes considerable displacement and, to a lesser extent, elastic deformation due not only to respiratory but also to cardiac motion. Although breathing is cyclic, its critical endpoint—that is, the
AJR:204, February 2015 287
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al. Fig. 1—Schema of operations. Registration of other phases to portal venous phase results in postregistered series. Portal venous phase dataset is unchanged. Postregistered subtracted series generated by workstation uses postregistered unenhanced phase as mask image.
Preregistered series
Postregistered series
Unenhanced phase
Postregistered unenhanced phase
Registered to portal venous phase
Arterial phase
Postregistered arterial phase
Portal venous phase
Portal venous phase
Delayed phase
Postregistered delayed phase
Preregistered subtraction
Postregistered subtraction
Arterial phase
Portal venous phase
breath-hold—may vary from one acquisition to the next. This variation leads to variations in the position of the liver during examinations with repeated breath-holds and to subsequent misregistration on subtraction images [3]. As one would expect, the maximum respiratory variation in displacement has been shown to occur in the craniocaudal direction [4]. However, some anteroposterior motion and mediolateral motion also occur, and there is significant heterogeneity in the relative displacement of different liver segments during respiration [5]. This difference in relative displacements of different liver segments occurs because, in addition to inconsistencies in the depth of breath-holds,
Delayed phase
Arterial phase
cardiac pulsation (most pronounced in the left lobe) and elastic deformation add to the complexity of intersegmental hepatic movement [3]. Cardiac pulsation and elastic deformation result in additional in-plane misregistration for lesions near the dome and capsule of the liver where subtraction “band” artifacts are often seen [6]. Applying a simple purely z-axis (craniocaudal) motion correction by rigid transformation (slice shifting) may not offer a satisfactory solution to this perplexing problem of in-plane misregistration. The usefulness of subtraction images for characterizing small and subcapsular lesions has therefore been limited because of the resulting misregistration artifacts [1].
A
Portal venous phase
Delayed phase
Registration of hepatic motion is of clinical interest in multiphasic breath-hold examinations to reduce subtraction artifacts and improve lesion characterization [5]. Various image registration techniques have been described for multiphasic examinations of the liver [7]. The purpose of this study was to quantify the displacement of the liver between the different dynamic phases in a multiphasic examination, measure the improvement in displacement after registration using anatomic landmarks, and quantify the improvements in image quality and reader confidence for evaluating treated HCC. Intrahepatic cysts were used as anatomic landmarks, and registration was performed by a fully automated
B
Fig. 2—Measurement of subtraction band artifacts at right hepatic vein–inferior vena cava (IVC) junction in 43-year-old woman with treated hepatocellular carcinoma. A and B, Preregistered (A) and postregistered (B) arterial phase images show decrease in maximum thickness of band artifact for all four quadrants. Vertical dashed line passes through mid IVC, and horizontal dashed line passes 2 cm anterior to IVC. Greatest band thickness perpendicular to liver surface (solid lines) is measured in each quadrant.
288
AJR:204, February 2015
The local institutional review board approved this HIPAA-compliant retrospective study and waived the requirement for informed consent. The data from 25 multiphasic liver MRI studies, all performed to evaluate treated HCC in cirrhotic patients (20 men and five women; age range, 35–72 years) scanned between January 2010 and December 2011, were analyzed. These patients were chosen for this study because, in addition to treated HCC, they also had small easily seen intrahepatic cysts; a total of 32 T1 hypointense cysts conspicuous in all dynamic phases were used as intrahepatic markers to measure displacement without registration and with registration. Each patient’s treated HCC lesion (n = 25) was evaluated for lesion conspicuity on nonregistered and registered subtraction images.
MR Image Acquisition Multiphasic contrast-enhanced MRI was performed on a 1.5-T scanner (Achieva, Philips Healthcare) using a 16-channel torso phased-array coil. An axial 3D two-echo T1 fast-field echo modified Dixon (Philips Healthcare) sequence was used with the following parameters: TR/TE1, TE2, 6.0/1.8, 4.0; flip angle, 10°; FOV, 38–40 cm; matrix size, 204–320 × 192–208; and scanning time per dynamic phase, 15–18 seconds [8]. Imaging was performed before and after IV injection of 0.1 mL/ kg of gadoteridol (ProHance, Bracco Diagnostics) using an MRI-compatible power injector (Spectris, Medrad) in end-expiratory breath-hold (one dynamic phase per breath-hold). Contrast-enhanced scans were manually triggered after fluoroscopic bolus arrival detection at the aortic bifurcation with approximate acquisition start times of 30 seconds (arterial phase), 70 seconds (portal venous phase), and 5 minutes (delayed phase) after injection.
Image Registration and Subtraction The breath-hold 3D modified Dixon images were transferred to a U.S. Food and Drug Administration–approved dedicated workstation (CADstream liver workstation, version 5.3.1.218, Merge Healthcare). Registered base and subtraction images were created using the workstation’s fully automated 3D deformable registration algorithm (Fig. 1). This propriety software, according to its default settings, registers unenhanced, arterial, and equilibrium phase data to the portal venous phase. The registration algorithm incorporates a fully automated combination of rigid and isotropic elastic nonrigid transformation using featurebased detection and matching methods that are in-
5.0
Reader 2
Reader 3
4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0
Arterial Phase
Portal Venous Phase
Delayed Phase
Arterial Phase
Preregistered Series
Portal Venous Phase
Delayed Phase
Postregistered Series
A 6.0 Reader 2
Reader 3
5.0 4.0 3.0 2.0 1.0 0
Arterial Phase
Portal Venous Phase
Delayed Phase
Arterial Phase
Preregistered Series
Portal Venous Phase
Delayed Phase
Postregistered Series
B 7 6
Fig. 3—Analyses of image quality of pre- and postregistered subtraction images. A, Bar graph shows increase in subjective image quality of conspicuity of treated hepatocellular carcinoma after registration (1 = nondiagnostic, 5 = excellent image quality). HCC = hepatocellular carcinoma. B, Bar graph shows increase in subjective image quality of entire subtraction volume after registration (1 = nondiagnostic, 5 = excellent image quality). C, Bar graph shows decrease in mean width of subtraction band artifact after registration.
dependent of the image intensity and therefore not influenced by implicit signal intensity data, which is the actual variable we are most interested in measuring in hepatic MRI [9]. The technical details of this registration algorithm can be found in the following references: [7, 10].
Mean Width of Subtraction Band Artifact (mm)
Materials and Methods Subjects
Subjective Grade for Image Quality of Entire Subtraction Volume
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
nonrigid transformation algorithm on a commercially available liver workstation.
Subjective Grade for Conspicuity of Treated HCC
Image Registration for Subtractions in Liver DCE-MRI
5 4 3 2 1 0
Arterial Phase
Portal Venous Phase
Preregistered series
Delayed Phase
Postregistered series
C Subtracted datasets were created by subtracting the unenhanced (mask) images from each of the contrast-enhanced phases (arterial, portal venous, and delayed). Subtraction of the nonregistered datasets resulted in the data henceforth referred to as “preregistered subtraction” data. Subtraction of
AJR:204, February 2015 289
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al.
A
B
C
D
E
F
Fig. 4—45-year-old man with treated hepatocellular carcinoma HCC in segment VIII. Poor image quality affects image registration and lesion depiction. A and B, Preregistered axial T1-weighted arterial (A) and delayed (B) images show blurring of liver margins and lesion margins (arrow). C and D, Postregistered axial arterial (C) and delayed (D) images show accentuation of blurring of lesion margins (arrow). Readers rated (side-by-side image degradation) preregistered source images as better than postregistered source images for depiction of HCC. Postregistration source image degradation occurred in five of 25 patients. E and F, Preregistered (E) and postregistered (F) subtracted arterial phase images. Distortion and blurring (arrow) are more pronounced in postregistered subtracted arterial phase image than in preregistered subtracted arterial phase image. Reduction of subtracted image quality (i.e., of lesion conspicuity and image quality of volume) after registration occurred for this patient only.
the registered datasets—henceforth referred to as “postregistered subtraction”—was performed after source image registration.
ments in craniocaudal, anteroposterior, and mediolateral directions were also evaluated for each image set. In livers with multiple cysts, the average cyst displacement was used for the analysis.
Displacement Measurements Intrahepatic subcentimeter T1 hypointense cysts detectable on all phases were chosen as discrete intrahepatic landmarks and were used to calculate hepatic displacement during multiphasic scanning. The center voxel location of each T1 hypointense focus was determined with respect to x-, y-, and z-axis coordinates (mediolateral, anteroposterior, and craniocaudal, respectively) on each pre- and postregistered dynamic image set by one blinded radiologist who had 5 years’ experience at the time of the study (reader 1). Euclidean displacement (E) in the unenhanced (U), arterial (A), and delayed (D) phases was calculated using the following formula: EU,A,D = {(XP – XU,A,D)2 + (YP – YU,A,D)2 + (ZP – ZU,A,D)2} where X, Y, and Z are the x-, y-, and z-axis coordinates in 3D; and the subscripts U, A, P, and D denote unenhanced, arterial, portal venous, and delayed phases, respectively. Single-axis displace-
290
Image Evaluation Randomized subjective analysis—Patient-identifying information was removed from the image sets, and the image sets were sent to the hospital PACS for quality analysis. These datasets were presented in a random manner to two blinded radiologists (reader 2 with 5 years’ experience and reader 3 with 4 years’ experience) who were unaware of the registration status. They independently analyzed the 25 treated HCCs on both pre- and postregistered subtraction images and assessed these images in two ways. First, the readers evaluated the subtracted images for diagnostic confidence in treated HCC depiction and enhancement pattern (HCC lesion conspicuity); henceforth, we refer to this parameter as “lesion conspicuity.” The readers assessed lesion conspicuity using the following 5-point scale: 1, nondiagnostic, lesion not visualized; 2, lesion barely visualized with enhancement fea-
tures not assessed because of severe artifacts; 3, lesion visualized with uncertain enhancement features; 4, lesion visualized with sufficiently discernible enhancement pattern and minimal artifacts or slight blurring of the edges; and 5, lesion visualized with well-characterized enhancement pattern and sharp margins. Grades 4 and 5 were considered acceptable image quality. Second, the readers characterized their overall level of confidence in subtraction quality by taking into account the entire subtraction volume as a whole, which we refer to henceforth as “image quality of the volume.” The image quality of the volume was graded using the following 5-point scale: 1, overall nondiagnostic image quality; 2, severe subtraction artifacts (maximum thickness of subtraction bands of > 5 mm or more than two thirds of either the liver margins or vascular markings are ill defined); 3, moderate subtraction artifacts (maximum thickness of subtraction bands of 2–5 mm or between one third and two thirds of either the liver margins or vascular markings are ill defined); 4, good overall image quality with minimal artifacts (maximum thickness of subtraction bands of < 2 mm or less than one third of either
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI the liver margins or vascular markings are ill defined); and 5, perfect overall subtraction quality without any artifacts. Grades 4 and 5 were considered satisfactory image quality. Side-by-side analysis—The maximum perpendicular thickness of the subtraction band artifact was measured by reader 1 in each of four quadrants on the liver surface at the level of the junction of the hepatic veins with the inferior vena cava in both the pre- and postregistered subtraction images (Fig. 2). The cysts, whose true enhancement nature was known (i.e., none), were chosen as targets to assess the burden of potential false-positive enhancement on the characterization of focal liver lesions. Focal false-positive enhancement in the subtracted images due to misregistration of the T1 hypointense cysts was measured by reader 1 in a binary fashion (yes or no). The preregistered source images were used as a reference standard to confirm or refute any findings of enhancement in these foci. Finally, to assess for potential source image degradation due to the image registration process, two radiologists (readers 2 and 3) reviewed the pre- and postregistered source images (blinded as to which) side-by-side to compare HCC lesion conspicuity and margin definition. The following question about the preregistered (A) and postregistered (B) series was asked: On which image set is the HCC lesion best seen (image series A or image series B), or were both image series equally good?”
Statistical Analysis The displacement of T1 hypointense foci and the width of the subtraction band artifact were analyzed using the two-way paired Student t test. The overall diagnostic confidence and image artifacts (lesion conspicuity and image quality of the volume) were compared between pre- and postregistered subtraction images using the Wilcoxon signed rank test. Evaluation of reader preference for preversus postregistered source images in terms of image degradation was done by Monte Carlo simulation of multinomial data. Interobserver agreement was assessed by Cohen’s kappa statistics.
Results Lesions in Study Population Two patients had three liver cysts used to calculate displacement, three patients had two cysts, and all others had a solitary liver cyst. Twenty-five treated HCCs (average size, 2.8 cm; range, 1.5–8.9 cm) were analyzed (one HCC per patient). These lesions were located in segments II (n = 2), III (n = 2), IVa (n = 4), V (n = 1), VI (n = 4), VII (n = 6), and VIII (n = 6). Thirteen treated HCCs were intrinsically hyperintense on T1weighted imaging, and 12 were hypointense on T1-weighted imaging. Seventeen of 25 HCCs (68%) were subcapsular in location, and eight (32%) lesions were central in location. Fifteen HCCs (60%) were located in
TABLE 1: Measured Displacements of Anatomic Landmarks Mean (SD) [Range] Displacement of Anatomic Landmark (mm) Displacement Measurement Method and Imaging Phase
Preregistered Subtraction Images
Postregistered Subtraction Images
p
Euclidian displacement Unenhanced
4.0 (2.6) [1.1–11.9]
2.4 (1.4) [0.1–5.3]
0.006
Arterial
3.2 (3.5) [0.0–18.0]
1.6 (1.4) [0.0–6.3]
0.004
Delayed
4.6 (3.7) [1.1–14.9]
1.3 (1.0) [0.0–5.1]
< 0.001
Unenhanced
2.5 (2.2) [0.0–8.8]
1.1 (1.6) [0.0–4.0]
< 0.001
Arterial
2.2 (2.7) [0.0–14.0]
0.8 (1.2) [0.0–4.0]
0.002
Delayed
2.9 (3.0) [0.0–12.0]
0.5 (1.0) [0.0–4.0]
< 0.001
Unenhanced
1.3 (1.4) [0.0–0.6]
1.0 (0.8) [0.0–4.0]
0.16
Arterial
0.9 (0.8) [0.0–3.3]
0.5 (0.4) [0.0–2.1]
0.02
Delayed
1.6 (1.2) [0.0–4.3]
0.6 (0.6) [0.0–2.7]
< 0.001
Unenhanced
1.5 (1.8) [0.0–10.1]
1.0 (0.9) [0.0–3.3]
0.11
Arterial
1.5 (2.1) [0.0–11.2]
0.8 (0.9) [0.0–4.3]
0.02
Delayed
1.6 (2.7) [0.0–13.3]
0.6 (0.6) [0.0–1.9]
0.04
Craniocaudal displacement
Mediolateral displacement
Anteroposterior displacement
the same segment as the cyst that was used as the marker for motion evaluation. The average distance of the cyst from the HCC was 3.9 cm (range, 1.7–10.1 cm). Residual disease was detected in two of these treated HCCs. Motion Assessment The mean displacements of the anatomic landmarks (i.e., the cysts) are listed in Table 1. The average Euclidian displacement was significantly lower after registration for all three phases (pre- vs postregistered measurements: unenhanced, 4.0 vs 2.4 mm, respectively; arterial, 3.2 vs 1.6 mm; delayed, 4.6 vs 1.3 mm). Craniocaudal displacement was greatest: The average craniocaudal displacement over the three phases was 2.6 mm preregistered and 0.8 mm postregistered. The improvement in craniocaudal displacement was significant for each of the three phases. The average displacement of all three phases in the anteroposterior and mediolateral directions were 1.5 and 1.3 mm preregistered, decreasing to 0.8 and 0.7 mm postregistered. The decrease in anteroposterior and mediolateral displacement after registration was significant for the arterial and delayed phases. Randomized Subjective Analysis Evaluation of the subtracted image sets showed there was a statistically significant increase in lesion conspicuity after image registration (average pre- vs postregistered grade, 3.4 and 4.4 respectively; p < 0.01) (Table 2 and Fig. 3A). The average percentage of satisfactory subtraction image sets (lesion conspicuity grades 4 and 5) increased with registration—from 38% of preregistered subtracted image sets to 90% after registration (Table 2 and Fig. 3A). The global subtraction artifact level (image quality of the volume) also significantly improved in postregistered subtractions (3.3 vs 4.6 for pre- and postregistered subtractions, respectively; p < 0.01). Furthermore, the average percentage of satisfactory subtraction image sets (image quality of volume grades 4 and 5) increased with registration—from 34% of preregistered subtracted image sets to 95% after registration (Table 2 and Fig. 3B). As can be seen from Table 2, the kappa values for the two readers range from 0.6 to 0.8, reflecting good to excellent agreement between the two readers. Side-by-Side Analysis The average width of the subtraction band artifact decreased with registration from 5.3 to
AJR:204, February 2015 291
Sundarakumar et al.
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
TABLE 2: Evaluation of Subtracted Images
Parameter Assessed and Imaging Phase
Preregistered Subtraction Images
κ
Postregistered Subtraction Images
κ
HCC lesion conspicuity, mean grade [no. of cases with grade of 1, 2, 3, 4, 5, respectively] Arterial phase
0.8a
0.7
Reader 2
3.4 [0, 0, 15, 9, 1]
4.4 [0, 1, 1, 11, 12]
Reader 3
3.4 [0, 0, 14, 10, 1]
4.4 [0, 1, 0, 12, 12]
Portal venous phase
0.6a
0.7
Reader 2
3.5 [0, 0, 14, 9, 2]
4.5 [0, 1, 1, 7, 16]
Reader 3
3.3 [0, 0, 16, 8, 1]
4.4 [0, 1, 2, 7, 15]
Delayed phase
0.6a
0.6
Reader 2
3.4 [0, 0, 15, 9, 1]
4.5 [0, 1, 3, 3, 18]
Reader 3
3.1 [0, 0, 19, 5, 1]
4.5 [0, 1, 2, 4, 18]
3.4
4.4
Average
a
Image quality of volume, mean grade [no. of cases with grade of 1, 2, 3, 4, 5, respectively] Arterial phase
0.6a
0.7
Reader 2
3.1 [0, 3, 16, 5, 1]
Reader 3
3.1 [0, 3, 14, 7, 1]
Portal venous phase
4.4 [0, 1, 0, 10, 14] 4.4 [0, 1, 0, 12, 12] 0.6a
0.7
Reader 2
3.5 [0, 0, 13, 10, 2]
4.8 [0, 1, 0, 2, 22]
Reader 3
3.4 [0, 0, 15, 10, 0]
4.7 [0, 1, 0, 4, 20]
Delayed phase
0.7a
0.8
Reader 2
3.4 [0, 2, 15, 5, 3]
4.7 [0, 1, 0, 5, 19]
Reader 3
3.3 [0, 2, 16, 5, 2]
4.7 [0, 1, 1, 3, 20]
3.3
4.6
Arterial phase
5.3 (3.7)
2.4 (2.1)
Portal venous phase
6.1 (6.6)
2.6 (1.7)
Delayed phase
5.2 (4.9)
2.8 (3.0)
Average
5.5 (4.6)
2.6 (2.2)
10/32
0/32
Average
a
Subtraction band width (mm), mean (SD)
False-positive enhancement, no. of cysts/total no. of cysts
a
ap < 0.01.
2.4 mm in the arterial phase subtractions, from 6.1 to 2.6 mm in the portal venous phase subtractions, and from 5.2 to 2.8 mm in the delayed phase subtractions (p < 0.01) (Table 2 and Fig. 3C). However, in one case the postregistered subtraction dataset was worse than the preregistered subtraction dataset (Fig. 4). In the evaluation for the potential effects of image degradation on lesion conspicuity between pre- and postregistered source images, reader 2 rated the preregistered source images as better in five cases, worse in two cases, and equal in 18 cases. Reader 3 had almost identical results, rating the preregistered source images better in
292
four cases, worse in three cases, and equal in 18 cases (κ = 0.3). This trend toward source image degradation after registration did not reach statistical significance (p > 0.07). Ten of 32 cysts (31%) showed false-positive enhancement due to misregistration on the preregistered subtracted images. No false-positive enhancement due to cyst misregistration was seen in the postregistered subtraction datasets (Fig. 5). Discussion The results of this study revealed that considerable hepatic displacement occurred be-
tween breath-holds in all patients and, as expected, was greatest in the craniocaudal direction. These results almost certainly reflect the predominately craniocaudal inconsistencies in diaphragmatic (and therefore liver) positioning during repeated breathholds. Investigators noted in a previous study that quiet breathing results in craniocaudal displacement ranging from 10 to 26 mm [11]. Our measurements of displacement were considerably less, averaging 2.6 mm (maximum, 14.0 mm) and provides good estimates of variations of hepatic positions between multiple expiratory breath-holds in clinical abdominal MRI. The mean displacement was greatest on the delayed phase. This greater displacement likely results from the lengthy delay (≈ 5 minutes) between the earlier phases (relatively rapid back-to-back unenhanced, arterial, portal venous phases) and the delayed phase, at which time patients may not remember the “feel” of that same breath-hold or may have moved. After registration, the mean displacement decreased in all three directions, with average Euclidian displacement decreasing from 3.9 to 1.8 mm; average craniocaudal displacement, from 2.6 to 0.8 mm; average mediolateral displacement, from 1.3 to 0.7 mm; and average anteroposterior displacement, from 1.5 to 0.8 mm. The reductions in Euclidian and craniocaudal displacements were statistically significant in all three dynamic phases. In one extreme instance, we noted a maximum craniocaudal displacement of 14.0 mm (Euclidian displacement, 18.0 mm) in the arterial phase that after registration corrected to 4 mm (Euclidian displacement, 6.25 mm). Larger displacements, particularly those confined to the craniocaudal axis could, in principle, be registered by a purely rigid transformation. Unlike craniocaudal translation, however, the mediolateral and anteroposterior displacements (which are similar in magnitude) more likely reflect elastic deformation and therefore depend predominantly on nonrigid transformation for motion correction. Using the rigid and isotropic elastic nonrigid algorithm under investigation, we found that both anteroposterior and mediolateral displacements were also reduced by registration, achieving significance for anteroposterior and mediolateral displacements in the arterial and delayed phases. Registration of subtraction images led to considerable improvement in subjective HCC lesion conspicuity. There was also a significant decrease in subtraction band ar-
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI
A
B
C
D
Fig. 5—55-year-old woman with treated hepatocellular carcinoma (HCC) in segment IVa. False-positive enhancement of cyst was seen on preregistered subtraction image. A, Axial T2-weighted image shows cyst in segment VII (arrowhead); this cyst was used to measure hepatic motion. B, Axial unenhanced T1-weighted image shows heterogeneous hyperintensity in ablated cavity (arrow) and same hypointense cyst in segment VII (arrowhead). C, Axial arterial phase preregistered subtracted image (overall image quality of entire subtraction volume, grade 3) shows nodular and heterogeneous enhancement in periphery (dashed arrow) and uncertain enhancement within ablation cavity (solid arrow) (lesion conspicuity, grade 3). False-positive enhancement (arrowhead) due to misregistration is seen in segment VII cyst. D, Postregistered subtraction image in arterial phase at same level as C (overall image quality of entire subtraction volume, grade 4) shows linear enhancement limited to ablation margins and no enhancement within ablated cavity (solid arrow) (lesion conspicuity, grade 5). There is no misregistration-related enhancement (i.e., falsepositive enhancement) in segment VII cyst (arrowhead). Significant reduction in band artifact is also noted, particularly along inferior border of liver (dashed arrow).
tifacts along the liver margin after registration, allowing improved assessment of the subcapsular regions of liver, especially in the dome and inferior margins. This decrease in subtraction band artifacts is particularly advantageous for characterizing subcapsular lesions (Fig. 6). Similarly, the overall subtraction artifacts (image quality of the volume) significantly decreased within the liver because of the associated decreased misregistration of intrahepatic structures such as vessels. This decrease in artifacts may in part explain the favorable trend toward improvement in lesion conspicuity; the percentage of subtraction images rated satisfactory before registration (38%; average total displacement, 3.9 mm) more than doubled after registration to 90% (average total displace-
ment, 1.8 mm) (Table 2). There was only one case in which the subtraction image quality (both lesion conspicuity and image quality of the volume) was worse after registration (Fig. 4). The decrease in subtraction image quality in the latter case was likely a result of patient motion during acquisition, which caused blurring of the hepatic margins and internal features and led to increased artifacts in the preregistered images. In this case, registration based on compromised edge definition caused further deterioration of the image quality due to the smoothing filters used in the algorithm. In all other cases, registration was successful in mitigating the misregistration in the subtracted images. Yu et al. [1] observed that for preregistered subtracted images, the enhancement
features of lesions larger than 2 cm could always be reliably characterized, whereas the enhancement features could be reliably characterized in only 51% of lesions smaller than 2 cm. With the improvement in imaging technology, the minimum size of a nodule for which the enhancement criteria for diagnosing HCC on CT or MRI are applicable has decreased from 2 to 1 cm [12, 13]. Therefore, as characterizable lesions become smaller, we anticipate that the deleterious effects of misregistration artifacts in subtraction imaging may become more pronounced. In this study, 10 HCC lesions were equal to or smaller than 2 cm. Both readers characterized images from postregistered subtracted datasets of eight of these lesions (80%) as being of acceptable diagnostic quality—that is, without significant
AJR:204, February 2015 293
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al.
A
B
C
D
Fig. 6—60-year-old woman with treated hepatocellular carcinoma (HCC). Presence of band artifact affects characterization of subcapsular lesions. A and B, Preregistered subtracted arterial (A) and portal venous (B) images show misregistration of perinephric fat on segment VI (overall image quality of entire subtraction volume, grade 3) causing false-positive (in venous phase) enhancement in treated HCC (solid arrow, B) and obscuring small cyst (dashed arrow, B) in same segment (lesion conspicuity, grade 3). C and D, No band artifacts are noted in postregistered subtracted arterial (C) and portal venous (D) images (overall image quality of entire subtraction volume, grade 5). Thin linear enhancement (solid arrow) is noted around treated HCC (lesion conspicuity, grade 5) in arterial and portal venous phases and no enhancement is noted in cyst (dashed arrow, D). Note emergence of coherent artifacts in postregistered images because feature-based image registration may also boost depiction of artifacts (e.g., motion) present in source images.
artifacts (lesion conspicuity grade 4 or 5). The conspicuity of all of these lesions was assessed as grade 3 on preregistered subtraction series. Thus, in eight of 10 lesions that were equal to or smaller 2 cm, the image quality was improved from clinically nonsatisfactory (lesion conspicuity grade 3 or lower) to clinically satisfactory (lesion conspicuity grade 4 or 5). Although the focus of this study was not the detection and full characterization of HCC lesions, this improved image quality potentially opens new frontiers for more accurate characterization of small (i.e., 1–2 cm) lesions; this need to characterize small lesions arises in part from the use
294
of the management-based Liver Imaging Reporting and Data System (LI-RADS) classification [14]. In the previously mentioned study, Yu et al. [1] also found that the quality of the subtraction images was significantly better for images of central lesions than for images of subcapsular lesions, the latter of which they observed were often obscured by band artifacts. In our study, although a separate analysis based on the location and size of the lesion was not performed, 17 of 25 (68%) HCC lesions were subcapsular, and most of the subcapsular HCC lesions were well characterized. The observations of both readers
showed that the lesion conspicuity grade after registration improved from 3 to 4 or 5 for 14 of the 17 subcapsular lesions, improved from 4 to 5 in two lesions, and remained grade 5 in one lesion because of the significant reduction in subtraction artifacts in the subcapsular region of the liver (Fig. 6). This decrease in artifacts may in part explain the overall trend toward significant improvement in lesion depiction (lesion conspicuity) on postregistered subtraction images, with the limitations mentioned earlier at least partly overcome by registration. Moreover, any false-positive enhancement of the cysts on the subtraction images resolved after regis-
AJR:204, February 2015
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Image Registration for Subtractions in Liver DCE-MRI tration, further suggesting the potential for improvement in the evaluation of small hepatic lesions [3]. When compared with landmark- or surface-based registration, the registration method used in this study is robust for a highly mobile and deformable organ such as the liver. The feature-based registration methods used are based on similarities between corresponding points in the images. This technique is not affected by signal intensity differences, which occur during various phases of multiphasic scanning because of contrast enhancement. Some earlier enhancement-driven methods have used iterative registration generated by fitting a predetermined model of contrast pharmacokinetics to the datasets. This approach requires complex models and is biased toward the model itself and therefore may not truly reflect the enhancement pattern in all cases [7]. However, 3D deformable registration, as evaluated in this study, can in principle alter or degrade the appearance of lesions. Although the data show that subtraction image quality is improved by registration, it is important to also consider whether lesions may be degraded by the deformation that may occur during registration (image degradation). In this small evaluation of only treated HCCs, reader 2 rated five preregistered source image datasets as superior, two as inferior, and 18 as equal to the postregistered image datasets. Similarly, reader 3 rated four preregistered source image datasets as superior, three as inferior, and 18 as equal to the postregistered image datasets. These observations had poor interobserver agreement and were not statistically significant. Some reduction in image quality was likely caused by the inherent effects of smoothing, which occur during registration and change the appearance of the lesions. Although the effects of smoothing were too subtle to detect in 18 of 25 patients, it would be pertinent to ultimately assess the true dimensions, edge sharpness, and contrast of the lesions in the preregistered images alongside the postregistered images to investigate any potential discrepancy caused by image distortion after registration. This assessment was beyond the scope of this study, so this issue remains an important consideration if lesion diagnosis were to be made exclusively from the registered source images and bears further investigation. We should also note that although registration corrects the position variation occurring at the time interval between mul-
tiphasic scans, the registration is particularly prone to the deleterious effects of motion occurring during the scan itself. This motion is likely caused by motion artifacts in the source images that impair the feature-based matching process occurring in registration. In this study, most source images were devoid of significant artifacts (high overall image quality); thus, the subtraction algorithm had a higher likelihood of success. Although not evaluated in this study, a dedicated workstation approach to liver evaluation offers simultaneous display of anatomically synchronized multiphasic datasets despite wide variations in the actual positioning between them. For example, the craniocaudal displacement in preregistered images averaged 2.6 mm with a maximum of 14.0 mm. This amount of displacement means that 2-mm-thick axial images depicting a lesion could be potentially seven slices apart on different phases. After registration, with an average craniocaudal displacement of 0.8 mm and a maximum of 4.4 mm, this disparity would be at a maximum of two slices. This smaller disparity facilitates automatic linking and scrolling ability and allows a simplified analysis of enhancement kinetics (e.g., an ROI cursor placed in one postregistered phase is better aligned with the other phases). Dedicated liver workstation interfaces can also include functions such as apparent diffusion coefficient mapping, contrast kinetics, and volumetric analysis in addition to signal intensity characterization. The potential scope of improvement in workflow underscores its importance in a busy hepatic imaging clinic. Although the time required to transfer the study to a workstation, execute image registration, and transfer the postregistered images back to the PACS and its influence on workflow were not systematically analyzed (a function of the retrospective nature of this project), our datasets were exported from a PACS and processed on the workstation in approximately 5–7 minutes per case. Note that this process can occur in the background while the radiologist attends to other matters, and the images can then be analyzed at any time. This study has several limitations. First, the sample size was relatively small for achieving a large-scale validation of the robustness of registration and for justifying the added resources (e.g., time, equipment) required. However, this study was designed to measure the performance of automated registration in the delineation of treated HCC
and determine whether there is potential benefit to registered subtraction images, both of which were done. Second, the displacement of a small number of discrete focal points (cysts) within the liver was evaluated as a simplified estimate of linear translation. The use of this strategy for our study means that rotational and nonrigid deformations were not directly measured; therefore, these results do not give us insights into the degree of elastic deformation that is occurring. Third, because of significant intersegmental heterogeneity in hepatic displacement during respiration, the cysts used as internal reference points for displacement may not truly reflect the displacement of the HCC if they were not situated in near proximity. Fourth, there was no objective means by which to measure the effects of registration-associated smoothing and distortion on image quality. Instead, we subjectively assessed for any HCC degradation by blinded side-by-side comparison of nonsubtracted images, and most cases showed no artifacts to compromise HCC lesion evaluation. As we previously stated, however, this strategy needs to be investigated further before diagnosis is performed exclusively from registered datasets. Finally, the time required to transfer the study to the workstation, execute image registration, and transfer the postregistered images back to the PACS and the influence of this time investment on workflow were not systematically analyzed (a function of the retrospective nature of this project). However, depending on server and workstation speed, our datasets were exported and processed on the workstation with an acceptable time penalty (≈ 5–7 minutes per case). This process can occur in the background while the radiologist attends to other matters, and the images can then be analyzed at any time. In conclusion, this study quantified liver displacement between sequential expiratory breath-holds in multiphasic MRI and showed significant reductions in these displacements after automated 3D image registration. A commercially available, fully automated nonrigid registration algorithm significantly reduced subtraction artifacts and improved diagnostic confidence in the assessment of treated HCC by correcting image misregistration due to inconsistent liver positioning during sequential breath-holds. This capability has potential implications for the detection and characterization of small (≤ 2 cm) and subcapsular lesions, which are traditionally difficult to evaluate with subtraction because of artifacts. These
AJR:204, February 2015 295
Downloaded from www.ajronline.org by NYU Langone Med Ctr-Sch of Med on 05/31/15 from IP address 128.122.253.212. Copyright ARRS. For personal use only; all rights reserved
Sundarakumar et al. preliminary results suggest registration may be helpful in routine liver MRI, particularly for follow-up of HCCs after ablation. Further work is required to evaluate the full impact of registration on lesion characterization. References 1. Yu JS, Kim YH, Rofsky NM. Dynamic subtraction magnetic resonance imaging of cirrhotic liver: assessment of high signal intensity lesions on nonenhanced T1-weighted images. J Comput Assist Tomogr 2005; 29:51–58 2. Kuehl H, Rosenbaum-Krumme S, Veit-Haibach P, et al. Impact of whole-body imaging on treatment decision to radio-frequency ablation in patients with malignant liver tumors: comparison of [18F] fluorodeoxyglucose-PET/computed tomography, PET and computed tomography. Nucl Med Commun 2008; 29:599–606 3. Yu JS, Rofsky NM. Dynamic subtraction MR imaging of the liver: advantages and pitfalls. AJR 2003; 180:1351–1357 4. Clifford MA, Banovac F, Levy E, Cleary K. As-
sessment of hepatic motion secondary to respiration for computer assisted interventions. Comput Aided Surg 2002; 7:291–299 5. Srimathveeravalli G, Leger J, Ezell P, Maybody M, Gutta N, Solomon SB. A study of porcine liver motion during respiration for improving targeting in image-guided needle placements. Int J Comput Assist Radiol Surg 2013; 8:15–27 6. Tatli S, Acar M, Tuncali K, Sadow CA, Morrison PR, Silverman SG. MRI assessment of percutaneous ablation of liver tumors: Value of subtraction images. J Magn Reson Imaging 2013; 37:407–413 7. Melbourne A, Atkinson D, White MJ, Collins D, Leach M, Hawkes D. Registration of dynamic contrast-enhanced MRI using a progressive principal component registration (PPCR). Phys Med Biol 2007; 52:5147–5156 8. Eggers H, Brendel B, Duijndam A, Herigault G. Dual-echo Dixon imaging with flexible choice of echo times. Magn Reson Med 2011; 65:96–107 9. Melbourne A, Atkinson D, Hawkes D. Influence of organ motion and contrast enhancement on image registration. Med Image Comput Comput As-
sist Interv 2008; 11(pt 2):948–955 10. Rajaraman S, Rodriguez JJ, Graff C, et al. Automated registration of sequential breath-hold dynamic contrast-enhanced MR images: a comparison of three techniques. Magn Reson Imaging 2011; 29:668–682 11. Suramo I, Päivänsalo M, Myllylä V. Cranio-caudal movements of the liver, pancreas and kidneys in respiration. Acta Radiol Diagn (Stockh) 1984; 25:129–131 12. Forner A, Vilana R, Ayuso C, et al. Diagnosis of hepatic nodules 20 mm or smaller in cirrhosis: Prospective validation of the noninvasive diagnostic criteria for hepatocellular carcinoma. Hepatology 2008; 47:97–104 13. Song do S, Bae SH. Changes of guidelines diagnosing hepatocellular carcinoma during the last tenyear period. Clin Mol Hepatol 2012; 18:258–267 14. Purysko AS, Remer EM, Coppa CP, Leão Filho HM, Thupili CR, Veniero JC. LI-RADS: a casebased review of the new categorization of liver findings in patients with end-stage liver disease. RadioGraphics 2012; 32:1977–1995
F O R YO U R I N F O R M AT I O N
Unique customized medical search engine service from ARRS! ARRS GoldMiner® is a keyword- and concept-driven search engine that provides instant access to radiologic images published in peer-reviewed journals. ARRS members earn 0.5 CME credits for each self-directed search. For more information, visit goldminer.arrs.org.
296
AJR:204, February 2015
