A three-dimensional head-and-neck phantom for validation of multimodality deformable image registration for adaptive radiotherapy Kamal Singhrao, Neil Kirby, and Jean Pouliot Citation: Medical Physics 41, 121709 (2014); doi: 10.1118/1.4901523 View online: http://dx.doi.org/10.1118/1.4901523 View Table of Contents: http://scitation.aip.org/content/aapm/journal/medphys/41/12?ver=pdfcov Published by the American Association of Physicists in Medicine Articles you may be interested in Accuracy of surface registration compared to conventional volumetric registration in patient positioning for head-and-neck radiotherapy: A simulation study using patient data Med. Phys. 41, 121701 (2014); 10.1118/1.4898103 A virtual phantom library for the quantification of deformable image registration uncertainties in patients with cancers of the head and neck Med. Phys. 40, 111703 (2013); 10.1118/1.4823467 Site-specific deformable imaging registration algorithm selection using patient-based simulated deformations Med. Phys. 40, 041911 (2013); 10.1118/1.4793723 Tissue characterization using a phantom to validate four-dimensional tissue deformation Med. Phys. 39, 6065 (2012); 10.1118/1.4747528 A realistic deformable prostate phantom for multimodal imaging and needle-insertion procedures Med. Phys. 39, 2031 (2012); 10.1118/1.3692179
A three-dimensional head-and-neck phantom for validation of multimodality deformable image registration for adaptive radiotherapy Kamal Singhrao, Neil Kirby, and Jean Pouliota) Department of Radiation Oncology, University of California San Francisco, San Francisco, California 94143-1708
(Received 27 March 2014; revised 28 September 2014; accepted for publication 27 October 2014; published 19 November 2014) Purpose: To develop a three-dimensional (3D) deformable head-and-neck (H&N) phantom with realistic tissue contrast for both kilovoltage (kV) and megavoltage (MV) imaging modalities and use it to objectively evaluate deformable image registration (DIR) algorithms. Methods: The phantom represents H&N patient anatomy. It is constructed from thermoplastic, which becomes pliable in boiling water, and hardened epoxy resin. Using a system of additives, the Hounsfield unit (HU) values of these materials were tuned to mimic anatomy for both kV and MV imaging. The phantom opens along a sagittal midsection to reveal radiotransparent markers, which were used to characterize the phantom deformation. The deformed and undeformed phantoms were scanned with kV and MV imaging modalities. Additionally, a calibration curve was created to change the HUs of the MV scans to be similar to kV HUs, (MC). The extracted ground-truth deformation was then compared to the results of two commercially available DIR algorithms, from Velocity Medical Solutions and software. Results: The phantom produced a 3D deformation, representing neck flexion, with a magnitude of up to 8 mm and was able to represent tissue HUs for both kV and MV imaging modalities. The two tested deformation algorithms yielded vastly different results. For kV–kV registration, produced mean and maximum errors of 1.8 and 11.5 mm, respectively. These same numbers for Velocity were 2.4 and 7.1 mm, respectively. For MV–MV, kV–MV, and kV–MC Velocity produced similar mean and maximum error values. , however, produced gross errors for all three of these scenarios, with maximum errors ranging from 33.4 to 41.6 mm. Conclusions: The application of DIR across different imaging modalities is particularly difficult, due to differences in tissue HUs and the presence of imaging artifacts. For this reason, DIR algorithms must be validated specifically for this purpose. The developed H&N phantom is an effective tool for this purpose. C 2014 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4901523] Key words: deformable phantom, deformation verification, deformable image registration, multimodality imaging, mutual information 1. INTRODUCTION The potential clinical use of deformable image registration (DIR) for applications such as transferring anatomical contours and previously delivered dose has been of immense interest to researchers. However, before clinical implementation can be adopted, DIR algorithms must be subjected to validation tests, such as landmark tracking,1–7 contour comparison,4,6,8–11 and both simulated1,7,12–14 and physical phantom deformations.1–3,7,15,16 Physical phantoms were developed to provide a complete end-to-end evaluation of the image acquisition and DIR process. Identification of landmarks in phantom images is commonly used to measure deformation. However, if these landmarks are also visible to DIR algorithms, their use to evaluate DIR accuracy can yield a skewed error measurement. Our group previously reported on a two-dimensional (2D) pelvic phantom for the evaluation of DIR algorithms.16 The key features of this phantom were that it was based on real patient anatomy and deformation, it was constructed from rigid (bone) and deformable (soft tissue) materials that 121709-1
Med. Phys. 41 (12), December 2014
mimicked Hounsfield unit (HU) values for actual anatomy and that it utilized optical markers to measure deformation. The computed tomography (CT) images of this phantom are a good surrogate for that of a patient, meaning that the results obtained therewith are relevant for real patients. Furthermore, the optical markers used to measure deformation are invisible on the phantom CT images. Hence, DIR algorithms are not biased by the markers and the true deformation is known accurately. The pelvic phantom was used to evaluate several widely available DIR algorithms and showed a wide variety of accuracy, with the percentage of points with errors larger than 3 mm ranging from 3.0% to 27.2%.16 The head-and-neck (H&N) phantom presented here is an evolution of the previous phantom. It has all key features of the pelvic phantom but also has several improvements. The first improvement is the extension of the deformation evaluation from 2D to three-dimensions (3D). Second, a new thermoplastic material was introduced to adequately represent tissue contrast for both kilovoltage (kV) and megavolatge (MV) imaging, thus enabling evaluation of registrations for CBCT (kV and/or MV) (provisional patent application,
0094-2405/2014/41(12)/121709/7/$30.00
© 2014 Am. Assoc. Phys. Med.
121709-1
121709-2
Singhrao, Kirby, and Pouliot: 3D H&N phantom for DIR validation
Attorney Docket No.: 84850-871225-016500US). This ability is especially important to evaluate the application of DIR to on-board imaging for adaptive radiotherapy. Finally, the new design and fabrication process allow for a more realistic 3D representation of a patient. Although, the new phantom is not a fully 3D rendering of a patient, it deforms and is characterized in 3D. Thus, this phantom is a complete end-to-end DIR evaluation tool that can be used within and across kV and MV imaging modalities.
2. METHODS AND MATERIALS 2.A. Phantom design and setup
Figure 1 displays the design for the phantom, which represents the central sagittal slice of the H&N anatomy. It opens along the sagittal midsection into two separate halves, which key into each other to ensure that both halves deform mutually. These keys are visible in the brain, which creates contrast in an otherwise homogeneous region. Since there is contrast in the brain for real patients, some contrast is an improvement over a homogeneous region. However, the keys might also create more contrast than that for a real patient and potentially skew
121709-2
DIR accuracy in this region. Each half of the phantom consists of a 2D slice containing anatomical structures and with a thickness of 16.5 mm. Beyond this, there is a 19.1 mm thick bolus representing the lateral edges of a patient. The anatomical details for the phantom and the bolus were modeled after the sagittal midplane of an actual H&N patient [see Fig. 1(e)], selected from the institution’s anonymized database. This model patient image was acquired on a Siemens CT scanner, with a voxel resolution of 0.98 × 0.98 × 3 mm. The opening midsection represented the measurable deformation plane and had a grid of 743 radiotransparent markers, with 5 mm spacing, painted on each half to physically characterize the magnitude of any deformation. An initial model of the phantom was milled from wood with computer-numerical-controlled (CNC) machining. Then, a silicone mold was made of this model, into which the doped solid thermoplastic polyurethane was cast. The heat deflection temperature of the solid polyurethane was 70 ◦C, enabling it to be warped in boiling water and for it to become fixed at room temperature. The base HU value of the solid polyurethane was measured to be 989 and 1021 HU for kV and MV CT images, respectively. The HU value of the urethane was tuned upward to that of muscle, brain, and spine tissues by the
F. 1. (Top row) (a) The different shades represent different tissue types. The locations of the spherical keys are shown as black (b) and white (c) disks on the phantom surface. The grid of radiotransparent markers [shown as black dots on (c) and (d)] is painted on the deformation plane. (d) An optical image of phantom half opened along the deformation plane. (Bottom) (e) The sagittal patient kVCT image, upon which the phantom was based. The corresponding kVCT (f) and MVCBCT (g) scans of phantom. Some of the spherical keys are visible in these images, which create some contrast in an otherwise fairly homogeneous region. (h) An optical image of the assembled phantom. Medical Physics, Vol. 41, No. 12, December 2014
121709-3
Singhrao, Kirby, and Pouliot: 3D H&N phantom for DIR validation
addition of aluminum trichloride powder and downward to that of fat via the addition of glass microspheres. To represent bony anatomy, high temperature epoxy with a heat deflection temperature of 135 ◦C, doped with calcium carbonate, was used. The materials were cast to fit as inserts inside a tissue characterization phantom and scanned under kV and MV CT modalities. Then, the additive levels were optimized for resemblance to the tissue inserts of a CIRS electron density phantom (Model 062). To generate the deformation, the forehead and sternum were held in place while pressure was applied to the back of the neck along the PA direction [see right arrow, Fig. 1(d)]. This caused the head to flex backward [curved arrow, Fig. 1(d)] and also to the side, thus simulating a slightly different head position on another treatment day. To measure the deformation, information was combined from optical and CT images. 2D coordinates of the radiotransparent markers were extracted using an optical camera via in-house software. These optical images were then rigidly registered to the sagittal midplane of phantom CT image to transfer these measurements to DICOM coordinates. The errors introduced by this transfer dominated the uncertainty in these coordinates and were estimated to be 0.25 and 0.4 mm in the anterior/posterior and superior/inferior directions, respectively. As a result of the design of the phantom, there is a thin air gap (
A three-dimensional head-and-neck phantom for validation of multimodality deformable image registration for adaptive radiotherapy Kamal Singhrao, Neil Kirby, and Jean Pouliota) Department of Radiation Oncology, University of California San Francisco, San Francisco, California 94143-1708
(Received 27 March 2014; revised 28 September 2014; accepted for publication 27 October 2014; published 19 November 2014) Purpose: To develop a three-dimensional (3D) deformable head-and-neck (H&N) phantom with realistic tissue contrast for both kilovoltage (kV) and megavoltage (MV) imaging modalities and use it to objectively evaluate deformable image registration (DIR) algorithms. Methods: The phantom represents H&N patient anatomy. It is constructed from thermoplastic, which becomes pliable in boiling water, and hardened epoxy resin. Using a system of additives, the Hounsfield unit (HU) values of these materials were tuned to mimic anatomy for both kV and MV imaging. The phantom opens along a sagittal midsection to reveal radiotransparent markers, which were used to characterize the phantom deformation. The deformed and undeformed phantoms were scanned with kV and MV imaging modalities. Additionally, a calibration curve was created to change the HUs of the MV scans to be similar to kV HUs, (MC). The extracted ground-truth deformation was then compared to the results of two commercially available DIR algorithms, from Velocity Medical Solutions and software. Results: The phantom produced a 3D deformation, representing neck flexion, with a magnitude of up to 8 mm and was able to represent tissue HUs for both kV and MV imaging modalities. The two tested deformation algorithms yielded vastly different results. For kV–kV registration, produced mean and maximum errors of 1.8 and 11.5 mm, respectively. These same numbers for Velocity were 2.4 and 7.1 mm, respectively. For MV–MV, kV–MV, and kV–MC Velocity produced similar mean and maximum error values. , however, produced gross errors for all three of these scenarios, with maximum errors ranging from 33.4 to 41.6 mm. Conclusions: The application of DIR across different imaging modalities is particularly difficult, due to differences in tissue HUs and the presence of imaging artifacts. For this reason, DIR algorithms must be validated specifically for this purpose. The developed H&N phantom is an effective tool for this purpose. C 2014 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4901523] Key words: deformable phantom, deformation verification, deformable image registration, multimodality imaging, mutual information 1. INTRODUCTION The potential clinical use of deformable image registration (DIR) for applications such as transferring anatomical contours and previously delivered dose has been of immense interest to researchers. However, before clinical implementation can be adopted, DIR algorithms must be subjected to validation tests, such as landmark tracking,1–7 contour comparison,4,6,8–11 and both simulated1,7,12–14 and physical phantom deformations.1–3,7,15,16 Physical phantoms were developed to provide a complete end-to-end evaluation of the image acquisition and DIR process. Identification of landmarks in phantom images is commonly used to measure deformation. However, if these landmarks are also visible to DIR algorithms, their use to evaluate DIR accuracy can yield a skewed error measurement. Our group previously reported on a two-dimensional (2D) pelvic phantom for the evaluation of DIR algorithms.16 The key features of this phantom were that it was based on real patient anatomy and deformation, it was constructed from rigid (bone) and deformable (soft tissue) materials that 121709-1
Med. Phys. 41 (12), December 2014
mimicked Hounsfield unit (HU) values for actual anatomy and that it utilized optical markers to measure deformation. The computed tomography (CT) images of this phantom are a good surrogate for that of a patient, meaning that the results obtained therewith are relevant for real patients. Furthermore, the optical markers used to measure deformation are invisible on the phantom CT images. Hence, DIR algorithms are not biased by the markers and the true deformation is known accurately. The pelvic phantom was used to evaluate several widely available DIR algorithms and showed a wide variety of accuracy, with the percentage of points with errors larger than 3 mm ranging from 3.0% to 27.2%.16 The head-and-neck (H&N) phantom presented here is an evolution of the previous phantom. It has all key features of the pelvic phantom but also has several improvements. The first improvement is the extension of the deformation evaluation from 2D to three-dimensions (3D). Second, a new thermoplastic material was introduced to adequately represent tissue contrast for both kilovoltage (kV) and megavolatge (MV) imaging, thus enabling evaluation of registrations for CBCT (kV and/or MV) (provisional patent application,
0094-2405/2014/41(12)/121709/7/$30.00
© 2014 Am. Assoc. Phys. Med.
121709-1
121709-2
Singhrao, Kirby, and Pouliot: 3D H&N phantom for DIR validation
Attorney Docket No.: 84850-871225-016500US). This ability is especially important to evaluate the application of DIR to on-board imaging for adaptive radiotherapy. Finally, the new design and fabrication process allow for a more realistic 3D representation of a patient. Although, the new phantom is not a fully 3D rendering of a patient, it deforms and is characterized in 3D. Thus, this phantom is a complete end-to-end DIR evaluation tool that can be used within and across kV and MV imaging modalities.
2. METHODS AND MATERIALS 2.A. Phantom design and setup
Figure 1 displays the design for the phantom, which represents the central sagittal slice of the H&N anatomy. It opens along the sagittal midsection into two separate halves, which key into each other to ensure that both halves deform mutually. These keys are visible in the brain, which creates contrast in an otherwise homogeneous region. Since there is contrast in the brain for real patients, some contrast is an improvement over a homogeneous region. However, the keys might also create more contrast than that for a real patient and potentially skew
121709-2
DIR accuracy in this region. Each half of the phantom consists of a 2D slice containing anatomical structures and with a thickness of 16.5 mm. Beyond this, there is a 19.1 mm thick bolus representing the lateral edges of a patient. The anatomical details for the phantom and the bolus were modeled after the sagittal midplane of an actual H&N patient [see Fig. 1(e)], selected from the institution’s anonymized database. This model patient image was acquired on a Siemens CT scanner, with a voxel resolution of 0.98 × 0.98 × 3 mm. The opening midsection represented the measurable deformation plane and had a grid of 743 radiotransparent markers, with 5 mm spacing, painted on each half to physically characterize the magnitude of any deformation. An initial model of the phantom was milled from wood with computer-numerical-controlled (CNC) machining. Then, a silicone mold was made of this model, into which the doped solid thermoplastic polyurethane was cast. The heat deflection temperature of the solid polyurethane was 70 ◦C, enabling it to be warped in boiling water and for it to become fixed at room temperature. The base HU value of the solid polyurethane was measured to be 989 and 1021 HU for kV and MV CT images, respectively. The HU value of the urethane was tuned upward to that of muscle, brain, and spine tissues by the
F. 1. (Top row) (a) The different shades represent different tissue types. The locations of the spherical keys are shown as black (b) and white (c) disks on the phantom surface. The grid of radiotransparent markers [shown as black dots on (c) and (d)] is painted on the deformation plane. (d) An optical image of phantom half opened along the deformation plane. (Bottom) (e) The sagittal patient kVCT image, upon which the phantom was based. The corresponding kVCT (f) and MVCBCT (g) scans of phantom. Some of the spherical keys are visible in these images, which create some contrast in an otherwise fairly homogeneous region. (h) An optical image of the assembled phantom. Medical Physics, Vol. 41, No. 12, December 2014
121709-3
Singhrao, Kirby, and Pouliot: 3D H&N phantom for DIR validation
addition of aluminum trichloride powder and downward to that of fat via the addition of glass microspheres. To represent bony anatomy, high temperature epoxy with a heat deflection temperature of 135 ◦C, doped with calcium carbonate, was used. The materials were cast to fit as inserts inside a tissue characterization phantom and scanned under kV and MV CT modalities. Then, the additive levels were optimized for resemblance to the tissue inserts of a CIRS electron density phantom (Model 062). To generate the deformation, the forehead and sternum were held in place while pressure was applied to the back of the neck along the PA direction [see right arrow, Fig. 1(d)]. This caused the head to flex backward [curved arrow, Fig. 1(d)] and also to the side, thus simulating a slightly different head position on another treatment day. To measure the deformation, information was combined from optical and CT images. 2D coordinates of the radiotransparent markers were extracted using an optical camera via in-house software. These optical images were then rigidly registered to the sagittal midplane of phantom CT image to transfer these measurements to DICOM coordinates. The errors introduced by this transfer dominated the uncertainty in these coordinates and were estimated to be 0.25 and 0.4 mm in the anterior/posterior and superior/inferior directions, respectively. As a result of the design of the phantom, there is a thin air gap (
