This copy is for personal use only. To order printed copies, contact [email protected]
Joseph Delic, MD Lea M. Alhilali, MD2 Marion A. Hughes, MD Serter Gumus, MD Saeed Fakhran, MD
Purpose:
To determine the performance of Shannon entropy (SE) as a diagnostic tool in patients with mild traumatic brain injury (mTBI) with posttraumatic migraines (PTMs) and those without PTMs on the basis of analysis of fractional anisotropy (FA) maps.
Materials and Methods:
The institutional review board approved this retrospective study, with waiver of informed consent. FA maps were obtained and neurocognitive testing was performed in 74 patients with mTBI (57 with PTM, 17 without PTM). FA maps were obtained in 22 healthy control subjects and in 20 control patients with migraine headaches. Mean FA and SE were extracted from total brain FA histograms and were compared between patients with mTBI and control subjects and between patients with and those without PTM. Mean FA and SE were correlated with clinical variables and were used to determine the areas under the receiver operating characteristic curve (AUCs) and likelihood ratios for mTBI and development of PTM.
Results:
Patients with mTBI had significantly lower SE (P , .001) and trended toward lower mean FA (P = .07) compared with control subjects. SE inversely correlated with time to recovery (TTR) (r = 20.272, P = .02). Patients with mTBI with PTM had significantly lower SE (P , .001) but not mean FA (P = .15) than did other patients with mTBI. SE provided better discrimination between patients with mTBI and control subjects than mean FA (AUC = 0.92; P = .01), as well as better discrimination between patients with mTBI with PTM and those without PTM (AUC = 0.85; P , .001). SE of less than 0.751 resulted in a 16.1 increased likelihood of having experienced mTBI and a 3.2 increased likelihood of developing PTM.
Conclusion:
SE more accurately reveals mTBI than mean FA, more accurately reveals those patients with mTBI who develop PTM, and inversely correlates with TTR. RSNA, 2016
q 1
From the Department of Radiology, Division of Neuroradiology, University of Pittsburgh Medical Center, 200 Lothrop St, PUH 2nd Floor, Suite 201 East Wing, Pittsburgh, PA 15213. Received June 24, 2015; revision requested August 4; revision received September 16; accepted September 29; final version accepted October 7. Address correspondence to L.A. (e-mail: [email protected]). Current address: Barrow Neurological Institute, Division of Neuroradiology, 350 W. Thomas Road, Phoenix, AZ 85013. 2
RSNA, 2016
q
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
1
Original Research n Neuroradiology
White Matter Injuries in Mild Traumatic Brain Injury and Posttraumatic Migraines: Diffusion Entropy Analysis1
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
M
ild traumatic brain injury (mTBI), often referred to as “concussion,” affects approximately 3.8 million people in the United States annually. Although these injuries are categorized as “mild,” their direct and indirect costs reach nearly $17 billion each year (1,2). Morbidity is substantial, with up to 50% of patients experiencing persistent difficulty at 1 month and 15%–25% experiencing difficulty at 1 year (3). Among patients with mTBI who have protracted symptoms, posttraumatic headache is common, with as many as 90% of patients experiencing headache symptoms (4). Posttraumatic migraines (PTMs) have especially high morbidity. PTMs are associated with lower neurocognitive test scores, aggravation of other posttraumatic symptoms, and a protracted recovery (5). Although the pathophysiology of PTM remains unclear, migraines outside of
Advances in Knowledge nn Shannon entropy (SE) analysis of fractional anisotropy (FA) histograms performed better as a diagnostic test to differentiate between patients with mild traumatic brain injury (mTBI) and control subjects than did mean FA (areas under the receiver operating characteristic curve [AUCs], 0.92 and 0.73, respectively; P = .01) and inversely correlated with time to recovery (r = 20.272, P = .02). nn SE performed better than mean FA in determining which patients with mTBI developed posttraumatic migraines (PTMs) (AUCs, 0.85 and 0.52, respectively; P , .001). nn SE of less than 0.750 was approximately 77% sensitive and 95% specific for the presence of mTBI and was 81% sensitive and 75% specific for PTM; patients with SE of less than 0.750 were approximately 16 times more likely to have experienced mTBI and were three times more likely to develop PTM. 2
trauma have been associated with focal white matter abnormalities (6). Given this association of nontraumatic migraines with white matter findings, we sought to determine if white matter injuries also play a role in PTMs. Advanced magnetic resonance (MR) imaging techniques with diffusion-tensor (DT) imaging hold promise in better depicting white matter injuries that are not apparent at conventional imaging (7– 10). DT imaging evaluation of focal white matter injuries after mTBI has been performed with both voxelwise and regionof-interest techniques (7–11); however, limitations to these approaches exist. Difficulty in analyzing crossing white matter tracts, interobserver variability in placing regions of interest, and decreased power as a result of the need to correct for multiple comparisons may affect results (12). In contrast to voxelwise or regional techniques, whole-brain histogram analysis does not rely on evaluation of individual white matter tracts, requires no manual analysis, and is ideally suited for pathologic conditions that are multifocal or diffuse, such as mTBI (7). Whole-brain histograms have demonstrated globally decreased mean fractional anisotropy (FA) in patients with mTBI relative to that in control subjects, and mean FA derived from histograms has correlated with injury severity along the entire spectrum of traumatic brain injury, from mild to severe (13,14). Unfortunately, all measures of central tendency, such as mean FA, are inherently limited by effects of voxel
Implications for Patient Care nn SE can be an effective data-mining tool to identify patients who have experienced a true concussive injury as well as to reveal those with more severe white matter injuries who may go on to develop PTMs. nn SE may provide a reproducible biomarker that can be calculated in automated fashion for diagnosis in patients with a higher likelihood of developing PTM and to predict recovery times in patients with mTBI.
Delic et al
averaging that may mask more subtle findings of tissue injury present in the FA histogram data. Thus, we looked to Shannon entropy (SE) to overcome these difficulties. SE is based on information theory and measures the complexity of a data set (15). The larger the amount of information or complexity contained in a data set, the more data points that are required to characterize the data set, and the higher the SE (16). This concept has been translated into the characterization of biologic tissues. Highly isotropic tissues, such as free water in cerebrospinal fluid, have a greater number of equal states. As a result, less data are required to describe this simple tissue and the SE is lower. However, in highly organized tissues, such as brain white matter, the high level of complexity requires more information to accurately describe the tissue, and the SE is higher (17). SE has proven superior to mean FA in revealing axonal remodeling after injury, is more sensitive to axonal density, and has been hypothesized to provide a more accurate reflection of axonal changes that occur after neurologic injury (18). We therefore hypothesize that SE can better reveal white
Published online before print 10.1148/radiol.2015151388 Content code: Radiology 2016; 000:1–8 Abbreviations: AUC = area under the ROC curve DT = diffusion tensor FA = fractional anisotropy ImPACT = Immediate Post-Concussion Assessment and Cognitive Testing mTBI = mild traumatic brain injury PTM = posttraumatic migraine ROC = receiver operating characteristic SE = Shannon entropy TTR = time to recovery Author contributions: Guarantors of integrity of entire study, L.M.A., S.F.; study concepts/study design or data acquisition or data analysis/ interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, all authors; clinical studies, J.D., L.M.A., S.G., S.F.; experimental studies, L.M.A.; statistical analysis, L.M.A.; and manuscript editing, all authors Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
matter injuries resulting from mTBI than traditional metrics of FA histograms alone. Thus, the purpose of this study was to determine the performance of SE as a diagnostic tool in patients with mTBI with PTMs and those without PTMs on the basis of analysis of FA maps.
Materials and Methods Subjects Our institutional review board approved this retrospective study, with waiver of informed consent. We searched our electronic medical record to identify MR imaging studies with DT imaging performed for mTBI. Radiology reports from January 1, 2006, to January 25, 2014, were searched by using the keywords “concussion,” “mild traumatic brain injury,” and “diffusion tensor imaging.” Inclusion criteria were age of 10–50 years, witnessed closed head trauma, no focal neurologic deficit, loss of consciousness of less than 1 minute, posttraumatic amnesia of less than 30 minutes, English language proficiency, and diagnosis of a concussion by a trained neuropsychologist. Exclusion criteria were prior neuropsychiatric illness (two patients), abnormal computed tomographic (CT) or conventional brain MR imaging findings (three patients), substance abuse (three patients), lack of DT imaging (four patients), lack of neurocognitive assessment (six patients), and a total symptom score of zero (three patients). Demographic data collected from the electronic medical record included age, sex, and history of migraines. Type of trauma (sports injury vs non-sports injury) and any history of a prior concussion were noted. The Immediate Post-Concussion Assessment and Cognitive Testing (ImPACT) system was utilized for serial computerized neurocognitive testing, including a total symptom score based on a seven-point Likert survey of 22 different postconcussion symptoms. Patients were classified as having migraine headaches on the basis of International Headache Society guidelines (19) following a postconcussion clinical
examination. Neuropsychologic and neurocognitive testing was performed by a neuropsychologist with more than 14 years of experience in treating patients with mTBI. Time to recovery (TTR) was defined as the point when the ImPACT total symptom score was zero or the patient stated that he or she was asymptomatic. Healthy control subjects were identified by searching the electronic medical record from January 1, 2006, to March 1, 2013, for MR imaging studies with the keywords “diffusion tensor,” “unremarkable,” and “normal.” Exclusion criteria were any history of prior traumatic brain injury or neurologic or psychiatric disease or any neuroradiologic finding at conventional MR imaging. Thirty-seven control studies were initially reviewed. Of these, eight were excluded for mTBI, seven for epilepsy, and one for abnormal T2 signal in the white matter. Reasons for the MR imaging examinations in the control subjects were headache (six subjects), prior infantile febrile seizure (three subjects), vertigo or dizziness (two subjects), question of an abnormality at an outside CT examination (three subjects), social dysfunction (two subjects), benign-appearing bony skull base lesion (two subjects), paresthesia (two subjects), and memory loss (one subject). Control patients with migraine were identified by searching the electronic medical record from January 1, 2006, to March 1, 2013, for MR imaging studies with the keywords “diffusion tensor” and “migraines.” Inclusion criteria were age less than 45 years, to better match the age distribution of mTBI, a history of migraines according to the International Headache Society guidelines (19), and DT imaging; exclusion criteria were any history of prior traumatic brain injury or neurologic or psychiatric disease apart from migraines or inadequate image quality.
DT and Conventional MR Imaging DT imaging was performed with a 1.5T system (Signa, GE Healthcare, Milwaukee, Wis) with a standard head coil. Despite the relatively extended time interval over which this study was performed, all patients underwent an
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
Delic et al
identical concussion DT imaging protocol: DT imaging was performed with a single-shot echo-planar sequence (repetition time msec/echo time msec, 4000/80; two excitations; section thickness, 5 mm; matrix, 128 3 128; and field of view, 260 mm). Diffusion gradients were set in 25 noncollinear directions by using two b values (0 and 1000 sec/mm2). The field of view ranged from 200 to 240 mm.
Whole-Brain Histogram Analysis The Functional Software Library software (FMRIB, version 1.1; Oxford Centre for Functional Magnetic Resonance Imaging of the Brain, University of Oxford, Oxford, England) was used for FA analysis (20). DT images were corrected for eddy current distortions, and the skull was “extracted” by using the FSL brain extraction tool, which applies a single binarized mask in diffusion space containing ones inside the brain and zeroes outside the brain (21). FA maps were then aligned into a common space by using nonlinear registration (22). Whole-brain histogram analysis was performed in an automated fashion for each patient’s aligned FA data by using 197 equal-width bins (Fig 1). SE was then automatically calculated as:
H (X 1 ) = 2 Σ p(x i ) log p (x i ),, xi ∈ k
where H is the entropy function, Xi is the FA values of the white matter skeleton, k is the number of bins, and p(xi) is the probability of a given FA value being repeated throughout (15). (Processing was performed by two neuroradiologists with 4 years of quantitative image analysis experience [L.M.A., S.F.] and a radiology resident with 1 year of experience in quantitative image analysis [J.D.]).
Statistical Analysis Comparison of demographic variables was performed by using a Fisher exact test or log-linear analysis. Comparison of the mean FA and SE of the FA values was performed with a two-sample t test or one-way analysis of variance (ANOVA) test for parametric distributions of the data and a Mann-Whitney U test 3
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Delic et al
Figure 1
Figure 1: Sample histograms for, A, a control subject and, B, a patient with mTBI. Despite the overall similar appearance of the curves, subjective differences in the complexity of the histogram shape are seen between the patient and the control subject.
or Kruskal-Wallis test for nonparametric distributions. For one-way ANOVA tests, a pairwise comparison was performed with a post-hoc Tukey test. For correlation of histogram parameters with continuous data, we used the Pearson correlation coefficient. Receiver operating characteristic (ROC) curves were created for mean FA and SE for the prediction of mTBI and PTM. Areas under the ROC curve (AUCs) for mean FA and SE and the prediction of mTBI and PTM were obtained and were interpreted according to Hosmer and Lemeshow (23) to show no discrimination (AUC = 0.5), poor discrimination (0.6 AUC 0.7), acceptable discrimination (0.7 , AUC 0.8), excellent discrimination (0.8 , AUC 0.9), and outstanding discrimination (AUC . 0.9). Comparisons of the AUCs for mean FA and SE were performed for both the prediction of mTBI and the presence of PTM. The optimal test value cutoff was determined by finding the point closest to the top-left part of the ROC plot for each ROC curve. Optimal test value cutoffs were then used to determine the likelihood ratios of the presence of mTBI and the development of PTM according to the method of Altman (24). All P values were two tailed, and P , .05 was considered to indicate a statistically significant difference. Analysis was performed by a physician with postgraduate statistics training (L.M.A.). 4
Figure 2
Figure 2: Histogram of the time to presentation for patients with mTBI shows that the majority of the patients (83%) presented within 2 months of injury. The remaining nine patients with mTBI presented more than 12 weeks after the initial injury.
Results Characteristics of the Subjects Seventy-four patients with mTBI (52 male patients; mean age, 18 years; range, 10–47 years) and 22 control subjects (10 male subjects; mean age, 18.8 years; range, 12–25 years) were included; 30 patients in the mTBI group (40%) had experienced a prior concussion. The median time from injury to
clinical presentation was 20 days (range, 0–506 days) (Fig 2). The most common trauma mechanisms were sports injury (43 patients [57%]) and motor vehicle accident (nine patients [12%]). The average neurocognitive test score percentiles were as follows: ImPACT total symptom score, 31.8 (range, 1–97); verbal memory score, 34.1 (range, 1–99); visual memory score, 28.63 (range, 1–97); processing speed score, 38.1 (range, 1–98); and reaction
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
time score, 35.6 (range, 1–97). The median TTR was 32 weeks (range, 2–252 weeks). Patients were not significantly different from healthy control subjects or control subjects with migraine in terms of age or sex (Table 1).
Comparison of Mean FA and SE by Using Whole-Brain Histogram Analysis Patients with mTBI trended toward having lower mean FA compared with healthy control subjects and control subjects with migraine and had significantly lower SE than either healthy control subjects or control subjects
with migraine (Table 1). No significant difference was seen between healthy control subjects and control subjects with migraine for either FA or SE at post-hoc analysis.
Correlation of FA and SE with Clinical Variables and Neurocognitive Testing Neither mean FA nor SE correlated with verbal memory, visual memory, reaction time, processing speed at neurocognitive testing (r absolute value range = 0.040–0.152, P value range = .19–.70), or symptom severity (r = 0.161 and 20.060, P = .17 and .61,
Table 1 Comparison of Demographic and Histogram Properties between Patients with mTBI and Control Subjects Parameter Mean age (y)* No. of male patients† Mean FA‡ Mean SE‡
Healthy Control Subjects (n = 22)
Migraine Control Subjects (n = 20)
Patients with mTBI (n = 74)
P Value
18.8 (6–44) 10 (48) 0.383 (0.379, 0.388) 0.808 (0.788, 0.847)
21.7 (16–43) 10 (50) 0.382 (0.365, 0.399) 0.795 (0.785, 0.805)
18.0 (10–47) 51 (68) 0.379 (0.371, 0.383) 0.737 (0.718, 0.763)
.11 .64 .07 ,.001
* Data in parentheses are the range. †
Data in parentheses are percentages.
‡
Data in parentheses are 95% confidence intervals.
Delic et al
respectively). SE was inversely correlated with TTR (r = 20.272, P = .02).
Comparison of Patients with mTBI with PTM and Those without PTM Patients with PTM were not significantly different from those without PTM in terms of age, sex, history of prior concussion, type of injury, or initial ImPACT total symptom score. Patients with PTM trended toward having a longer recovery time (P = .05) and had significantly lower SE than those without PTM (P = .002), but not mean FA (P = .15) (Table 2). ROC Curve Analysis Mean FA demonstrated acceptable performance in discriminating between patients with mTBI and control subjects (AUC = 0.73). However, SE demonstrated outstanding discrimination between patients with mTBI and control subjects (AUC = 0.92), which was significantly better than the performance of mean FA (P = .01). SE also demonstrated excellent discrimination in predicting the presence of PTM (AUC = 0.85), significantly better than the poor performance of mean FA (AUC = 0.52, P , .001). ROC curves
Table 2 Comparison of Clinical Factors, Mean FA, and SE between Patients with mTBI with PTM and Those without PTM Parameter Mean age (y)† No. of male patients‡ No. of patients with prior concussion‡ No. of patients with sports injury‡ Average ImPACT total symptom score percentile†§ Average verbal memory score percentile† Average visual memory score percentile† Average reaction time percentile† Average processing speed percentile† Median TTR (w)† Mean FA|| Mean SE||
Patients with PTM (n = 57)
Patients without PTM (n = 17)
P Value*
17.6 (10–38) 39 (67) 24 (41) 34 (59) 36.1 (1–97) 30.2 (1–99) 28.1 (1–97) 34.6 (1–95) 33.5 (1–98) 51.9 (1–252) 0.378 (0.366, 0.390) 0.732 (0.692, 0.772)
19.7 (12–47) 12 (71) 8 (47) 8 (47) 20.8 (1–74) 37.5 (7–92) 36.5 (1–88)|| 43.1 (1–94) 47.7 (1–94) 39.4 (3–194) 0.375 (0.367, 0.383) 0.757 (0.687, 0.827)
.26 ..99 .79 .56 .42 .01 .20 .13 .18 .05 .15 .002
* P values were two tailed and were calculated with the use of an unpaired t test for continuous variables and with a Fisher exact test for categorical variables. †
Data in parentheses are the range.
‡
Data in parentheses are percentages.
§
ImPACT scores are percentiles determined by normative data from baseline testing of more than 17 000 athletes as part of their pre-sport participation, with percentile information accounting for both sex and age.
||
Data are 95% confidence intervals.
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
5
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Figure 3
Figure 3: ROC curves for, A, mean FA and, B, SE show that mean FA provides acceptable discrimination for differentiating patients with mTBI from control subjects, with an AUC of 0.73, while SE provides outstanding discrimination (AUC = 0.92; P = .01). The difference becomes more pronounced in attempting to predict the presence of PTM, where, C, mean FA shows poor performance (AUC = 0.52) and, D, SE continues to demonstrate excellent discrimination (AUC = 0.85; P , .001).
are shown in Figure 3. The sensitivities and specificities for the optimality criterion for the mean FA and SE on the basis of their ROC curves are shown in Table 3.
Discussion SE analysis of whole-brain FA histograms demonstrated outstanding diagnostic performance for distinguishing patients who have suffered a true concussive injury from control subjects and was further able to help distinguish those patients with mTBI with PTM. Lower SE also correlated with increased recovery time in patients with mTBI. These findings may indicate 6
that SE can be used to reach a more accurate diagnosis in patients with true concussive injuries and provide a biomarker of recovery. Multiple previous studies have evaluated the use of histogram analysis in mTBI (13,14,16), with mixed results. These prior studies have focused primarily on mean FA, which can be problematic in pathologic conditions such as mTBI, where changes in DT imaging metrics can be bidirectional, reflecting the opposite effects of axonal injury and axonal swelling on FA. These bidirectional changes may negate each other in traditional histogram metrics. However, unlike traditional histogram properties, SE reflects the complexity
Delic et al
of a system and is able to account for bidirectional data. As a result, SE has been used in other clinical applications where there are opposite directions of change, such as electrocardiographic analysis (25,26). Similarly, in mTBI, a change in the absolute FA may be delayed by pseudonormalization from competing effects of the initial injury, remyelination, and swelling (7); however, SE will still increase during this time, as the white matter structure becomes more complex, even if no change in absolute FA is seen. As a result, SE may serve as a better biomarker of recovery, and possibly aid in recovery decisions regarding whether to return to sports or other high-risk activities. SE has been previously used in other disease processes that demonstrate a similar pattern of early elevated FA, late decreased FA, and pseudonormalization of the absolute FA value (27). In cervical myelopathy, early cord compression results in increased FA (27), with subsequent decreased FA as gliosis, demyelination, and extracellular edema replace initial axonal swelling (28), similar to the temporal changes seen in mTBI in our current study. However, despite the temporal heterogeneity, decreased SE within the cervical cord was able to consistently differentiate patients with spondylosis and myelopathy from those without myelopathy, even when there were no significant differences at conventional imaging or even in mean FA (29). Our current study demonstrates that SE can be used as an effective diagnostic tool in mTBI as well. SE histogram analysis has the potential to overcome not only the temporal variability in FA but also the variability seen in individual DT imaging measurements, because it is based on FA values in all voxels. Thus, variations in single tracts or voxels that may be confounding in comparisons of a single region are less significant when averaged over the entire brain. The SE histogram method also combines the strengths of previous voxelwise and region-of-interest techniques. Like voxelwise methods, it is hypothesis free, can be used to evaluate the entire brain in one analysis session, and is automated. However,
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Table 3 Optimality Criterion Values and Associated Sensitivities and Specificities for Mean FA and SE from ROC Curve Analysis Comparison Mean FA in patients with mTBI vs that in control subjects SE in patients with mTBI vs that in control subjects Mean FA in patients with mTBI with PTM vs that in patients with mTBI without PTM SE in patients with mTBI with PTM vs that in patients with mTBI without PTM
Optimality Criterion Value
Sensitivity (%)
Specificity (%)
Likelihood Ratio*
0.381
67 (54, 77)
81 (57, 94)
3.5 (1.4, 8.6)
0.751
77 (65, 86)
95 (74, 100)
16.1 (2.4, 109.7)
0.375
35 (23, 49)
59 (33, 81)
0.9 (0.8, 1.7)
0.750
81 (67, 90)
75 (43, 93)
3.2 (1.2, 8.7)
Note.—Data in parentheses are 95% confidence intervals. * Likelihood ratio was a conventional likelihood ratio, not weighted by prevalence.
like region-of-interest techniques, it preserves statistical power, without the need for multiple comparisons. Our study had several limitations, including its retrospective format and moderately sized cohort. A prospective study with a larger sample size would assist in corroborating these findings. In addition, our study included both patients who were believed clinically to warrant imaging and those with prior concussions. Thus, a selection bias may have existed toward more seriously injured patients who had substantial symptoms at presentation that warranted imaging. Arguably, although a bias may have existed, it was a bias toward the patients who would most benefit from imaging biomarkers. Additionally, as our study involved whole-brain analysis, no regional differences could be assessed. Nor could we assess the contribution of gray matter changes to the overall SE changes. Further regional analysis of SE changes should be performed to better localize the injuries detected at the overall histogram analysis. Our study results suggest that SE analysis of FA histograms may better reflect the white matter pathologic conditions underlying mTBI and postconcussive symptoms and may have a role as a diagnostic tool and prognostic biomarker in individual patients with concussion. This is a novel concept that has the potential to refine the clinical
treatment of individual patients with mTBI. In the future, SE analysis of DT images may be used to aid clinicians in more accurately diagnosing concussion, anticipating symptoms, and predicting prognosis in mTBI. Disclosures of Conflicts of Interest: J.D. disclosed no relevant relationships. L.M.A. disclosed no relevant relationships. M.A.H. disclosed no relevant relationships. S.G. disclosed no relevant relationships. S.F. disclosed no relevant relationships.
References 1. Finkelstein EA, Corso PS, Miller TR. The incidence and economic burden of injuries in the United States. New York, NY: Oxford University Press, 2006. 2. Centers for Disease Control and Prevention. Traumatic brain injury in the United States: emergency department visits, hospitalizations and deaths 2002-2006. Atlanta, Ga: Centers for Disease Control and Prevention, 2010. 3. Bazarian JJ, McClung J, Shah MN, Cheng YT, Flesher W, Kraus J. Mild traumatic brain injury in the United States, 1998-2000. Brain Inj 2005;19(2):85–91. 4. Kontos AP, Elbin RJ, Lau B, et al. Posttraumatic migraine as a predictor of recovery and cognitive impairment after sport-related concussion. Am J Sports Med 2013;41(7): 1497–1504. 5. Erickson JC. Treatment outcomes of chronic post-traumatic headaches after mild head trauma in US soldiers: an observational study. Headache 2011;51(6):932–944.
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
Delic et al
6. Yu D, Yuan K, Qin W, et al. Axonal loss of white matter in migraine without aura: a tract-based spatial statistics study. Cephalalgia 2013;33(1):34–42. 7. Niogi SN, Mukherjee P. Diffusion tensor imaging of mild traumatic brain injury. J Head Trauma Rehabil 2010;25(4):241–255. 8. Alhilali LM, Yaeger K, Collins M, Fakhran S. Detection of central white matter injury underlying vestibulopathy after mild traumatic brain injury. Radiology 2014;272(1):224–232. 9. Fakhran S, Yaeger K, Alhilali L. Symptomatic white matter changes in mild traumatic brain injury resemble pathologic features of early Alzheimer dementia. Radiology 2013;269(1):249–257. 10. Eierud C, Craddock RC, Fletcher S, et al. Neuroimaging after mild traumatic brain injury: Review and meta-analysis. Neuroimage Clin 2014;4:283–294. 11. Fakhran S, Yaeger K, Collins M, Alhilali L. Sex differences in white matter abnormalities after mild traumatic brain injury: localization and correlation with outcome. Radiology 2014;272(3):815–823. 12. Wintermark M, Sanelli PC, Anzai Y, Tsiouris AJ, Whitlow CT; American College of Radiology Head Injury Institute. Imaging evidence and recommendations for traumatic brain injury: advanced neuro- and neurovascular imaging techniques. AJNR Am J Neuroradiol 2015;36(2):E1–E11. 13. Lipton ML, Gellella E, Lo C, et al. Multifocal white matter ultrastructural abnormalities in mild traumatic brain injury with cognitive disability: a voxel-wise analysis of diffusion tensor imaging. J Neurotrauma 2008;25(11): 1335–1342. 14. Benson RR, Meda SA, Vasudevan S, et al. Global white matter analysis of diffusion tensor images is predictive of injury severity in traumatic brain injury. J Neurotrauma 2007;24(3):446–459. 15. Shannon CE. The mathematical theory of communication. 1963. MD Comput 1997;14(4): 306–317. 16. Inglese M, Makani S, Johnson G, et al. Diffuse axonal injury in mild traumatic brain injury: a diffusion tensor imaging study. J Neurosurg 2005;103(2):298–303. 17. Metwalli NS, LaConte SM, Hu XP. An in formation theoretic approach characterizing diffusion anisotropy in diffusion-weighted magnetic resonance images. Conf Proc IEEE Eng Med Biol Soc 2006;1:2260–2263. 18. Fozouni N, Chopp M, Nejad-Davarani SP, et al. Characterizing brain structures and remodeling after TBI based on information
7
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
content, diffusion entropy. PLoS One 2013; 8(10):e76343. 19. Headache Classification Subcommittee of the International Headache Society. The International Classification of Headache Disorders: 2nd edition. Cephalalgia 2004;24(Suppl 1): 9–160. 20. Smith SM, Jenkinson M, Woolrich MW, et al. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 2004;23(Suppl 1):S208– S219.
tion using free-form deformations: application to breast MR images. IEEE Trans Med Imaging 1999;18(8):712–721. 23. Hosmer DW, Lemeshow S. Applied logistic regression. 2nd ed. Hoboken, NJ: Wiley, 2000; 156–164. 24. Altman DG. Practical statistics for medical research. London, England: Chapman & Hall, 1990; 268–271, 352, 354, 417.
21. Smith SM. Fast robust automated brain extraction. Hum Brain Mapp 2002;17(3):143–155.
25. Zhou T, Lin D, Yang C, Wu X, Fang Z. Analysis of epicardial mapping electrogram of sustained atrial fibrillation based on Shannon entropy. Conf Proc IEEE Eng Med Biol Soc 2009;2009:3470–3472.
22. Rueckert D, Sonoda LI, Hayes C, Hill DL, Leach MO, Hawkes DJ. Nonrigid registra-
26. Alcaraz R, Rieta JJ. Wavelet bidomain sample entropy analysis to predict spontaneous
8
Delic et al
termination of atrial fibrillation. Physiol Meas 2008;29(1):65–80. 27. Guan X, Fan G, Wu X, et al. Diffusion tensor imaging studies of cervical spondylotic myelopathy: a systemic review and metaanalysis. PLoS One 2015;10(2):e0117707. 28. Baron EM, Young WF. Cervical spondylotic myelopathy: a brief review of its pathophysiology, clinical course, and diagnosis. Neurosurgery 2007;60(1 Supp1 1):S35–S41. 29. Cui JL, Wen CY, Hu Y, Li TH, Luk KD. Entropy-based analysis for diffusion anisotropy mapping of healthy and myelopathic spinal cord. Neuroimage 2011;54(3):2125–2131.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
Joseph Delic, MD Lea M. Alhilali, MD2 Marion A. Hughes, MD Serter Gumus, MD Saeed Fakhran, MD
Purpose:
To determine the performance of Shannon entropy (SE) as a diagnostic tool in patients with mild traumatic brain injury (mTBI) with posttraumatic migraines (PTMs) and those without PTMs on the basis of analysis of fractional anisotropy (FA) maps.
Materials and Methods:
The institutional review board approved this retrospective study, with waiver of informed consent. FA maps were obtained and neurocognitive testing was performed in 74 patients with mTBI (57 with PTM, 17 without PTM). FA maps were obtained in 22 healthy control subjects and in 20 control patients with migraine headaches. Mean FA and SE were extracted from total brain FA histograms and were compared between patients with mTBI and control subjects and between patients with and those without PTM. Mean FA and SE were correlated with clinical variables and were used to determine the areas under the receiver operating characteristic curve (AUCs) and likelihood ratios for mTBI and development of PTM.
Results:
Patients with mTBI had significantly lower SE (P , .001) and trended toward lower mean FA (P = .07) compared with control subjects. SE inversely correlated with time to recovery (TTR) (r = 20.272, P = .02). Patients with mTBI with PTM had significantly lower SE (P , .001) but not mean FA (P = .15) than did other patients with mTBI. SE provided better discrimination between patients with mTBI and control subjects than mean FA (AUC = 0.92; P = .01), as well as better discrimination between patients with mTBI with PTM and those without PTM (AUC = 0.85; P , .001). SE of less than 0.751 resulted in a 16.1 increased likelihood of having experienced mTBI and a 3.2 increased likelihood of developing PTM.
Conclusion:
SE more accurately reveals mTBI than mean FA, more accurately reveals those patients with mTBI who develop PTM, and inversely correlates with TTR. RSNA, 2016
q 1
From the Department of Radiology, Division of Neuroradiology, University of Pittsburgh Medical Center, 200 Lothrop St, PUH 2nd Floor, Suite 201 East Wing, Pittsburgh, PA 15213. Received June 24, 2015; revision requested August 4; revision received September 16; accepted September 29; final version accepted October 7. Address correspondence to L.A. (e-mail: [email protected]). Current address: Barrow Neurological Institute, Division of Neuroradiology, 350 W. Thomas Road, Phoenix, AZ 85013. 2
RSNA, 2016
q
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
1
Original Research n Neuroradiology
White Matter Injuries in Mild Traumatic Brain Injury and Posttraumatic Migraines: Diffusion Entropy Analysis1
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
M
ild traumatic brain injury (mTBI), often referred to as “concussion,” affects approximately 3.8 million people in the United States annually. Although these injuries are categorized as “mild,” their direct and indirect costs reach nearly $17 billion each year (1,2). Morbidity is substantial, with up to 50% of patients experiencing persistent difficulty at 1 month and 15%–25% experiencing difficulty at 1 year (3). Among patients with mTBI who have protracted symptoms, posttraumatic headache is common, with as many as 90% of patients experiencing headache symptoms (4). Posttraumatic migraines (PTMs) have especially high morbidity. PTMs are associated with lower neurocognitive test scores, aggravation of other posttraumatic symptoms, and a protracted recovery (5). Although the pathophysiology of PTM remains unclear, migraines outside of
Advances in Knowledge nn Shannon entropy (SE) analysis of fractional anisotropy (FA) histograms performed better as a diagnostic test to differentiate between patients with mild traumatic brain injury (mTBI) and control subjects than did mean FA (areas under the receiver operating characteristic curve [AUCs], 0.92 and 0.73, respectively; P = .01) and inversely correlated with time to recovery (r = 20.272, P = .02). nn SE performed better than mean FA in determining which patients with mTBI developed posttraumatic migraines (PTMs) (AUCs, 0.85 and 0.52, respectively; P , .001). nn SE of less than 0.750 was approximately 77% sensitive and 95% specific for the presence of mTBI and was 81% sensitive and 75% specific for PTM; patients with SE of less than 0.750 were approximately 16 times more likely to have experienced mTBI and were three times more likely to develop PTM. 2
trauma have been associated with focal white matter abnormalities (6). Given this association of nontraumatic migraines with white matter findings, we sought to determine if white matter injuries also play a role in PTMs. Advanced magnetic resonance (MR) imaging techniques with diffusion-tensor (DT) imaging hold promise in better depicting white matter injuries that are not apparent at conventional imaging (7– 10). DT imaging evaluation of focal white matter injuries after mTBI has been performed with both voxelwise and regionof-interest techniques (7–11); however, limitations to these approaches exist. Difficulty in analyzing crossing white matter tracts, interobserver variability in placing regions of interest, and decreased power as a result of the need to correct for multiple comparisons may affect results (12). In contrast to voxelwise or regional techniques, whole-brain histogram analysis does not rely on evaluation of individual white matter tracts, requires no manual analysis, and is ideally suited for pathologic conditions that are multifocal or diffuse, such as mTBI (7). Whole-brain histograms have demonstrated globally decreased mean fractional anisotropy (FA) in patients with mTBI relative to that in control subjects, and mean FA derived from histograms has correlated with injury severity along the entire spectrum of traumatic brain injury, from mild to severe (13,14). Unfortunately, all measures of central tendency, such as mean FA, are inherently limited by effects of voxel
Implications for Patient Care nn SE can be an effective data-mining tool to identify patients who have experienced a true concussive injury as well as to reveal those with more severe white matter injuries who may go on to develop PTMs. nn SE may provide a reproducible biomarker that can be calculated in automated fashion for diagnosis in patients with a higher likelihood of developing PTM and to predict recovery times in patients with mTBI.
Delic et al
averaging that may mask more subtle findings of tissue injury present in the FA histogram data. Thus, we looked to Shannon entropy (SE) to overcome these difficulties. SE is based on information theory and measures the complexity of a data set (15). The larger the amount of information or complexity contained in a data set, the more data points that are required to characterize the data set, and the higher the SE (16). This concept has been translated into the characterization of biologic tissues. Highly isotropic tissues, such as free water in cerebrospinal fluid, have a greater number of equal states. As a result, less data are required to describe this simple tissue and the SE is lower. However, in highly organized tissues, such as brain white matter, the high level of complexity requires more information to accurately describe the tissue, and the SE is higher (17). SE has proven superior to mean FA in revealing axonal remodeling after injury, is more sensitive to axonal density, and has been hypothesized to provide a more accurate reflection of axonal changes that occur after neurologic injury (18). We therefore hypothesize that SE can better reveal white
Published online before print 10.1148/radiol.2015151388 Content code: Radiology 2016; 000:1–8 Abbreviations: AUC = area under the ROC curve DT = diffusion tensor FA = fractional anisotropy ImPACT = Immediate Post-Concussion Assessment and Cognitive Testing mTBI = mild traumatic brain injury PTM = posttraumatic migraine ROC = receiver operating characteristic SE = Shannon entropy TTR = time to recovery Author contributions: Guarantors of integrity of entire study, L.M.A., S.F.; study concepts/study design or data acquisition or data analysis/ interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, all authors; clinical studies, J.D., L.M.A., S.G., S.F.; experimental studies, L.M.A.; statistical analysis, L.M.A.; and manuscript editing, all authors Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
matter injuries resulting from mTBI than traditional metrics of FA histograms alone. Thus, the purpose of this study was to determine the performance of SE as a diagnostic tool in patients with mTBI with PTMs and those without PTMs on the basis of analysis of FA maps.
Materials and Methods Subjects Our institutional review board approved this retrospective study, with waiver of informed consent. We searched our electronic medical record to identify MR imaging studies with DT imaging performed for mTBI. Radiology reports from January 1, 2006, to January 25, 2014, were searched by using the keywords “concussion,” “mild traumatic brain injury,” and “diffusion tensor imaging.” Inclusion criteria were age of 10–50 years, witnessed closed head trauma, no focal neurologic deficit, loss of consciousness of less than 1 minute, posttraumatic amnesia of less than 30 minutes, English language proficiency, and diagnosis of a concussion by a trained neuropsychologist. Exclusion criteria were prior neuropsychiatric illness (two patients), abnormal computed tomographic (CT) or conventional brain MR imaging findings (three patients), substance abuse (three patients), lack of DT imaging (four patients), lack of neurocognitive assessment (six patients), and a total symptom score of zero (three patients). Demographic data collected from the electronic medical record included age, sex, and history of migraines. Type of trauma (sports injury vs non-sports injury) and any history of a prior concussion were noted. The Immediate Post-Concussion Assessment and Cognitive Testing (ImPACT) system was utilized for serial computerized neurocognitive testing, including a total symptom score based on a seven-point Likert survey of 22 different postconcussion symptoms. Patients were classified as having migraine headaches on the basis of International Headache Society guidelines (19) following a postconcussion clinical
examination. Neuropsychologic and neurocognitive testing was performed by a neuropsychologist with more than 14 years of experience in treating patients with mTBI. Time to recovery (TTR) was defined as the point when the ImPACT total symptom score was zero or the patient stated that he or she was asymptomatic. Healthy control subjects were identified by searching the electronic medical record from January 1, 2006, to March 1, 2013, for MR imaging studies with the keywords “diffusion tensor,” “unremarkable,” and “normal.” Exclusion criteria were any history of prior traumatic brain injury or neurologic or psychiatric disease or any neuroradiologic finding at conventional MR imaging. Thirty-seven control studies were initially reviewed. Of these, eight were excluded for mTBI, seven for epilepsy, and one for abnormal T2 signal in the white matter. Reasons for the MR imaging examinations in the control subjects were headache (six subjects), prior infantile febrile seizure (three subjects), vertigo or dizziness (two subjects), question of an abnormality at an outside CT examination (three subjects), social dysfunction (two subjects), benign-appearing bony skull base lesion (two subjects), paresthesia (two subjects), and memory loss (one subject). Control patients with migraine were identified by searching the electronic medical record from January 1, 2006, to March 1, 2013, for MR imaging studies with the keywords “diffusion tensor” and “migraines.” Inclusion criteria were age less than 45 years, to better match the age distribution of mTBI, a history of migraines according to the International Headache Society guidelines (19), and DT imaging; exclusion criteria were any history of prior traumatic brain injury or neurologic or psychiatric disease apart from migraines or inadequate image quality.
DT and Conventional MR Imaging DT imaging was performed with a 1.5T system (Signa, GE Healthcare, Milwaukee, Wis) with a standard head coil. Despite the relatively extended time interval over which this study was performed, all patients underwent an
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
Delic et al
identical concussion DT imaging protocol: DT imaging was performed with a single-shot echo-planar sequence (repetition time msec/echo time msec, 4000/80; two excitations; section thickness, 5 mm; matrix, 128 3 128; and field of view, 260 mm). Diffusion gradients were set in 25 noncollinear directions by using two b values (0 and 1000 sec/mm2). The field of view ranged from 200 to 240 mm.
Whole-Brain Histogram Analysis The Functional Software Library software (FMRIB, version 1.1; Oxford Centre for Functional Magnetic Resonance Imaging of the Brain, University of Oxford, Oxford, England) was used for FA analysis (20). DT images were corrected for eddy current distortions, and the skull was “extracted” by using the FSL brain extraction tool, which applies a single binarized mask in diffusion space containing ones inside the brain and zeroes outside the brain (21). FA maps were then aligned into a common space by using nonlinear registration (22). Whole-brain histogram analysis was performed in an automated fashion for each patient’s aligned FA data by using 197 equal-width bins (Fig 1). SE was then automatically calculated as:
H (X 1 ) = 2 Σ p(x i ) log p (x i ),, xi ∈ k
where H is the entropy function, Xi is the FA values of the white matter skeleton, k is the number of bins, and p(xi) is the probability of a given FA value being repeated throughout (15). (Processing was performed by two neuroradiologists with 4 years of quantitative image analysis experience [L.M.A., S.F.] and a radiology resident with 1 year of experience in quantitative image analysis [J.D.]).
Statistical Analysis Comparison of demographic variables was performed by using a Fisher exact test or log-linear analysis. Comparison of the mean FA and SE of the FA values was performed with a two-sample t test or one-way analysis of variance (ANOVA) test for parametric distributions of the data and a Mann-Whitney U test 3
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Delic et al
Figure 1
Figure 1: Sample histograms for, A, a control subject and, B, a patient with mTBI. Despite the overall similar appearance of the curves, subjective differences in the complexity of the histogram shape are seen between the patient and the control subject.
or Kruskal-Wallis test for nonparametric distributions. For one-way ANOVA tests, a pairwise comparison was performed with a post-hoc Tukey test. For correlation of histogram parameters with continuous data, we used the Pearson correlation coefficient. Receiver operating characteristic (ROC) curves were created for mean FA and SE for the prediction of mTBI and PTM. Areas under the ROC curve (AUCs) for mean FA and SE and the prediction of mTBI and PTM were obtained and were interpreted according to Hosmer and Lemeshow (23) to show no discrimination (AUC = 0.5), poor discrimination (0.6 AUC 0.7), acceptable discrimination (0.7 , AUC 0.8), excellent discrimination (0.8 , AUC 0.9), and outstanding discrimination (AUC . 0.9). Comparisons of the AUCs for mean FA and SE were performed for both the prediction of mTBI and the presence of PTM. The optimal test value cutoff was determined by finding the point closest to the top-left part of the ROC plot for each ROC curve. Optimal test value cutoffs were then used to determine the likelihood ratios of the presence of mTBI and the development of PTM according to the method of Altman (24). All P values were two tailed, and P , .05 was considered to indicate a statistically significant difference. Analysis was performed by a physician with postgraduate statistics training (L.M.A.). 4
Figure 2
Figure 2: Histogram of the time to presentation for patients with mTBI shows that the majority of the patients (83%) presented within 2 months of injury. The remaining nine patients with mTBI presented more than 12 weeks after the initial injury.
Results Characteristics of the Subjects Seventy-four patients with mTBI (52 male patients; mean age, 18 years; range, 10–47 years) and 22 control subjects (10 male subjects; mean age, 18.8 years; range, 12–25 years) were included; 30 patients in the mTBI group (40%) had experienced a prior concussion. The median time from injury to
clinical presentation was 20 days (range, 0–506 days) (Fig 2). The most common trauma mechanisms were sports injury (43 patients [57%]) and motor vehicle accident (nine patients [12%]). The average neurocognitive test score percentiles were as follows: ImPACT total symptom score, 31.8 (range, 1–97); verbal memory score, 34.1 (range, 1–99); visual memory score, 28.63 (range, 1–97); processing speed score, 38.1 (range, 1–98); and reaction
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
time score, 35.6 (range, 1–97). The median TTR was 32 weeks (range, 2–252 weeks). Patients were not significantly different from healthy control subjects or control subjects with migraine in terms of age or sex (Table 1).
Comparison of Mean FA and SE by Using Whole-Brain Histogram Analysis Patients with mTBI trended toward having lower mean FA compared with healthy control subjects and control subjects with migraine and had significantly lower SE than either healthy control subjects or control subjects
with migraine (Table 1). No significant difference was seen between healthy control subjects and control subjects with migraine for either FA or SE at post-hoc analysis.
Correlation of FA and SE with Clinical Variables and Neurocognitive Testing Neither mean FA nor SE correlated with verbal memory, visual memory, reaction time, processing speed at neurocognitive testing (r absolute value range = 0.040–0.152, P value range = .19–.70), or symptom severity (r = 0.161 and 20.060, P = .17 and .61,
Table 1 Comparison of Demographic and Histogram Properties between Patients with mTBI and Control Subjects Parameter Mean age (y)* No. of male patients† Mean FA‡ Mean SE‡
Healthy Control Subjects (n = 22)
Migraine Control Subjects (n = 20)
Patients with mTBI (n = 74)
P Value
18.8 (6–44) 10 (48) 0.383 (0.379, 0.388) 0.808 (0.788, 0.847)
21.7 (16–43) 10 (50) 0.382 (0.365, 0.399) 0.795 (0.785, 0.805)
18.0 (10–47) 51 (68) 0.379 (0.371, 0.383) 0.737 (0.718, 0.763)
.11 .64 .07 ,.001
* Data in parentheses are the range. †
Data in parentheses are percentages.
‡
Data in parentheses are 95% confidence intervals.
Delic et al
respectively). SE was inversely correlated with TTR (r = 20.272, P = .02).
Comparison of Patients with mTBI with PTM and Those without PTM Patients with PTM were not significantly different from those without PTM in terms of age, sex, history of prior concussion, type of injury, or initial ImPACT total symptom score. Patients with PTM trended toward having a longer recovery time (P = .05) and had significantly lower SE than those without PTM (P = .002), but not mean FA (P = .15) (Table 2). ROC Curve Analysis Mean FA demonstrated acceptable performance in discriminating between patients with mTBI and control subjects (AUC = 0.73). However, SE demonstrated outstanding discrimination between patients with mTBI and control subjects (AUC = 0.92), which was significantly better than the performance of mean FA (P = .01). SE also demonstrated excellent discrimination in predicting the presence of PTM (AUC = 0.85), significantly better than the poor performance of mean FA (AUC = 0.52, P , .001). ROC curves
Table 2 Comparison of Clinical Factors, Mean FA, and SE between Patients with mTBI with PTM and Those without PTM Parameter Mean age (y)† No. of male patients‡ No. of patients with prior concussion‡ No. of patients with sports injury‡ Average ImPACT total symptom score percentile†§ Average verbal memory score percentile† Average visual memory score percentile† Average reaction time percentile† Average processing speed percentile† Median TTR (w)† Mean FA|| Mean SE||
Patients with PTM (n = 57)
Patients without PTM (n = 17)
P Value*
17.6 (10–38) 39 (67) 24 (41) 34 (59) 36.1 (1–97) 30.2 (1–99) 28.1 (1–97) 34.6 (1–95) 33.5 (1–98) 51.9 (1–252) 0.378 (0.366, 0.390) 0.732 (0.692, 0.772)
19.7 (12–47) 12 (71) 8 (47) 8 (47) 20.8 (1–74) 37.5 (7–92) 36.5 (1–88)|| 43.1 (1–94) 47.7 (1–94) 39.4 (3–194) 0.375 (0.367, 0.383) 0.757 (0.687, 0.827)
.26 ..99 .79 .56 .42 .01 .20 .13 .18 .05 .15 .002
* P values were two tailed and were calculated with the use of an unpaired t test for continuous variables and with a Fisher exact test for categorical variables. †
Data in parentheses are the range.
‡
Data in parentheses are percentages.
§
ImPACT scores are percentiles determined by normative data from baseline testing of more than 17 000 athletes as part of their pre-sport participation, with percentile information accounting for both sex and age.
||
Data are 95% confidence intervals.
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
5
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Figure 3
Figure 3: ROC curves for, A, mean FA and, B, SE show that mean FA provides acceptable discrimination for differentiating patients with mTBI from control subjects, with an AUC of 0.73, while SE provides outstanding discrimination (AUC = 0.92; P = .01). The difference becomes more pronounced in attempting to predict the presence of PTM, where, C, mean FA shows poor performance (AUC = 0.52) and, D, SE continues to demonstrate excellent discrimination (AUC = 0.85; P , .001).
are shown in Figure 3. The sensitivities and specificities for the optimality criterion for the mean FA and SE on the basis of their ROC curves are shown in Table 3.
Discussion SE analysis of whole-brain FA histograms demonstrated outstanding diagnostic performance for distinguishing patients who have suffered a true concussive injury from control subjects and was further able to help distinguish those patients with mTBI with PTM. Lower SE also correlated with increased recovery time in patients with mTBI. These findings may indicate 6
that SE can be used to reach a more accurate diagnosis in patients with true concussive injuries and provide a biomarker of recovery. Multiple previous studies have evaluated the use of histogram analysis in mTBI (13,14,16), with mixed results. These prior studies have focused primarily on mean FA, which can be problematic in pathologic conditions such as mTBI, where changes in DT imaging metrics can be bidirectional, reflecting the opposite effects of axonal injury and axonal swelling on FA. These bidirectional changes may negate each other in traditional histogram metrics. However, unlike traditional histogram properties, SE reflects the complexity
Delic et al
of a system and is able to account for bidirectional data. As a result, SE has been used in other clinical applications where there are opposite directions of change, such as electrocardiographic analysis (25,26). Similarly, in mTBI, a change in the absolute FA may be delayed by pseudonormalization from competing effects of the initial injury, remyelination, and swelling (7); however, SE will still increase during this time, as the white matter structure becomes more complex, even if no change in absolute FA is seen. As a result, SE may serve as a better biomarker of recovery, and possibly aid in recovery decisions regarding whether to return to sports or other high-risk activities. SE has been previously used in other disease processes that demonstrate a similar pattern of early elevated FA, late decreased FA, and pseudonormalization of the absolute FA value (27). In cervical myelopathy, early cord compression results in increased FA (27), with subsequent decreased FA as gliosis, demyelination, and extracellular edema replace initial axonal swelling (28), similar to the temporal changes seen in mTBI in our current study. However, despite the temporal heterogeneity, decreased SE within the cervical cord was able to consistently differentiate patients with spondylosis and myelopathy from those without myelopathy, even when there were no significant differences at conventional imaging or even in mean FA (29). Our current study demonstrates that SE can be used as an effective diagnostic tool in mTBI as well. SE histogram analysis has the potential to overcome not only the temporal variability in FA but also the variability seen in individual DT imaging measurements, because it is based on FA values in all voxels. Thus, variations in single tracts or voxels that may be confounding in comparisons of a single region are less significant when averaged over the entire brain. The SE histogram method also combines the strengths of previous voxelwise and region-of-interest techniques. Like voxelwise methods, it is hypothesis free, can be used to evaluate the entire brain in one analysis session, and is automated. However,
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
Table 3 Optimality Criterion Values and Associated Sensitivities and Specificities for Mean FA and SE from ROC Curve Analysis Comparison Mean FA in patients with mTBI vs that in control subjects SE in patients with mTBI vs that in control subjects Mean FA in patients with mTBI with PTM vs that in patients with mTBI without PTM SE in patients with mTBI with PTM vs that in patients with mTBI without PTM
Optimality Criterion Value
Sensitivity (%)
Specificity (%)
Likelihood Ratio*
0.381
67 (54, 77)
81 (57, 94)
3.5 (1.4, 8.6)
0.751
77 (65, 86)
95 (74, 100)
16.1 (2.4, 109.7)
0.375
35 (23, 49)
59 (33, 81)
0.9 (0.8, 1.7)
0.750
81 (67, 90)
75 (43, 93)
3.2 (1.2, 8.7)
Note.—Data in parentheses are 95% confidence intervals. * Likelihood ratio was a conventional likelihood ratio, not weighted by prevalence.
like region-of-interest techniques, it preserves statistical power, without the need for multiple comparisons. Our study had several limitations, including its retrospective format and moderately sized cohort. A prospective study with a larger sample size would assist in corroborating these findings. In addition, our study included both patients who were believed clinically to warrant imaging and those with prior concussions. Thus, a selection bias may have existed toward more seriously injured patients who had substantial symptoms at presentation that warranted imaging. Arguably, although a bias may have existed, it was a bias toward the patients who would most benefit from imaging biomarkers. Additionally, as our study involved whole-brain analysis, no regional differences could be assessed. Nor could we assess the contribution of gray matter changes to the overall SE changes. Further regional analysis of SE changes should be performed to better localize the injuries detected at the overall histogram analysis. Our study results suggest that SE analysis of FA histograms may better reflect the white matter pathologic conditions underlying mTBI and postconcussive symptoms and may have a role as a diagnostic tool and prognostic biomarker in individual patients with concussion. This is a novel concept that has the potential to refine the clinical
treatment of individual patients with mTBI. In the future, SE analysis of DT images may be used to aid clinicians in more accurately diagnosing concussion, anticipating symptoms, and predicting prognosis in mTBI. Disclosures of Conflicts of Interest: J.D. disclosed no relevant relationships. L.M.A. disclosed no relevant relationships. M.A.H. disclosed no relevant relationships. S.G. disclosed no relevant relationships. S.F. disclosed no relevant relationships.
References 1. Finkelstein EA, Corso PS, Miller TR. The incidence and economic burden of injuries in the United States. New York, NY: Oxford University Press, 2006. 2. Centers for Disease Control and Prevention. Traumatic brain injury in the United States: emergency department visits, hospitalizations and deaths 2002-2006. Atlanta, Ga: Centers for Disease Control and Prevention, 2010. 3. Bazarian JJ, McClung J, Shah MN, Cheng YT, Flesher W, Kraus J. Mild traumatic brain injury in the United States, 1998-2000. Brain Inj 2005;19(2):85–91. 4. Kontos AP, Elbin RJ, Lau B, et al. Posttraumatic migraine as a predictor of recovery and cognitive impairment after sport-related concussion. Am J Sports Med 2013;41(7): 1497–1504. 5. Erickson JC. Treatment outcomes of chronic post-traumatic headaches after mild head trauma in US soldiers: an observational study. Headache 2011;51(6):932–944.
Radiology: Volume 000: Number 0— 2016 n radiology.rsna.org
Delic et al
6. Yu D, Yuan K, Qin W, et al. Axonal loss of white matter in migraine without aura: a tract-based spatial statistics study. Cephalalgia 2013;33(1):34–42. 7. Niogi SN, Mukherjee P. Diffusion tensor imaging of mild traumatic brain injury. J Head Trauma Rehabil 2010;25(4):241–255. 8. Alhilali LM, Yaeger K, Collins M, Fakhran S. Detection of central white matter injury underlying vestibulopathy after mild traumatic brain injury. Radiology 2014;272(1):224–232. 9. Fakhran S, Yaeger K, Alhilali L. Symptomatic white matter changes in mild traumatic brain injury resemble pathologic features of early Alzheimer dementia. Radiology 2013;269(1):249–257. 10. Eierud C, Craddock RC, Fletcher S, et al. Neuroimaging after mild traumatic brain injury: Review and meta-analysis. Neuroimage Clin 2014;4:283–294. 11. Fakhran S, Yaeger K, Collins M, Alhilali L. Sex differences in white matter abnormalities after mild traumatic brain injury: localization and correlation with outcome. Radiology 2014;272(3):815–823. 12. Wintermark M, Sanelli PC, Anzai Y, Tsiouris AJ, Whitlow CT; American College of Radiology Head Injury Institute. Imaging evidence and recommendations for traumatic brain injury: advanced neuro- and neurovascular imaging techniques. AJNR Am J Neuroradiol 2015;36(2):E1–E11. 13. Lipton ML, Gellella E, Lo C, et al. Multifocal white matter ultrastructural abnormalities in mild traumatic brain injury with cognitive disability: a voxel-wise analysis of diffusion tensor imaging. J Neurotrauma 2008;25(11): 1335–1342. 14. Benson RR, Meda SA, Vasudevan S, et al. Global white matter analysis of diffusion tensor images is predictive of injury severity in traumatic brain injury. J Neurotrauma 2007;24(3):446–459. 15. Shannon CE. The mathematical theory of communication. 1963. MD Comput 1997;14(4): 306–317. 16. Inglese M, Makani S, Johnson G, et al. Diffuse axonal injury in mild traumatic brain injury: a diffusion tensor imaging study. J Neurosurg 2005;103(2):298–303. 17. Metwalli NS, LaConte SM, Hu XP. An in formation theoretic approach characterizing diffusion anisotropy in diffusion-weighted magnetic resonance images. Conf Proc IEEE Eng Med Biol Soc 2006;1:2260–2263. 18. Fozouni N, Chopp M, Nejad-Davarani SP, et al. Characterizing brain structures and remodeling after TBI based on information
7
NEURORADIOLOGY: Diffusion Entropy Analysis of White Matter Injuries in Mild Traumatic Brain Injury
content, diffusion entropy. PLoS One 2013; 8(10):e76343. 19. Headache Classification Subcommittee of the International Headache Society. The International Classification of Headache Disorders: 2nd edition. Cephalalgia 2004;24(Suppl 1): 9–160. 20. Smith SM, Jenkinson M, Woolrich MW, et al. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 2004;23(Suppl 1):S208– S219.
tion using free-form deformations: application to breast MR images. IEEE Trans Med Imaging 1999;18(8):712–721. 23. Hosmer DW, Lemeshow S. Applied logistic regression. 2nd ed. Hoboken, NJ: Wiley, 2000; 156–164. 24. Altman DG. Practical statistics for medical research. London, England: Chapman & Hall, 1990; 268–271, 352, 354, 417.
21. Smith SM. Fast robust automated brain extraction. Hum Brain Mapp 2002;17(3):143–155.
25. Zhou T, Lin D, Yang C, Wu X, Fang Z. Analysis of epicardial mapping electrogram of sustained atrial fibrillation based on Shannon entropy. Conf Proc IEEE Eng Med Biol Soc 2009;2009:3470–3472.
22. Rueckert D, Sonoda LI, Hayes C, Hill DL, Leach MO, Hawkes DJ. Nonrigid registra-
26. Alcaraz R, Rieta JJ. Wavelet bidomain sample entropy analysis to predict spontaneous
8
Delic et al
termination of atrial fibrillation. Physiol Meas 2008;29(1):65–80. 27. Guan X, Fan G, Wu X, et al. Diffusion tensor imaging studies of cervical spondylotic myelopathy: a systemic review and metaanalysis. PLoS One 2015;10(2):e0117707. 28. Baron EM, Young WF. Cervical spondylotic myelopathy: a brief review of its pathophysiology, clinical course, and diagnosis. Neurosurgery 2007;60(1 Supp1 1):S35–S41. 29. Cui JL, Wen CY, Hu Y, Li TH, Luk KD. Entropy-based analysis for diffusion anisotropy mapping of healthy and myelopathic spinal cord. Neuroimage 2011;54(3):2125–2131.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2016
