Neuroimaging biomarkers for predicting treatment response in schizophrenia based on alteration patterns of the whole brain white matter tracts
Jing-Ying Huang1, Yu-Jen Chen1, Chih-Min Liu2, Tzung-Jeng Hwang2,3, Yun-Chin Hsu1, Yu-Chun Lo1, Hai-Gwo Hwu2,3, and Wen-Yih Isaac Tseng1,3,4,5,6

1Institute of Medical Device and Imaging, National Taiwan University College of Medicine, Taipei, Taiwan, 2Department of Psychiatry, National Taiwan University Hospital, Taipei, Taiwan, 3Graduate Institute of Brain and Mind Sciences, National Taiwan University College of Medicine, Taipei, Taiwan, 4Institute of Biomedical Engineering, National Taiwan University, Taipei, Taiwan, 5Department of Medical Imaging, National Taiwan University Hospital, Taipei, Taiwan, 6Molecular Imaging Center, National Taiwan University, Taipei, Taiwan


This study aims to identify image biomarkers for schizophrenia patients in order to predict treatment responses on individual subject basis. We develop algorithm that can discriminate remission or non-remission in patients with schizophrenia based on the difference in microstructural integrity of the white matter tracts. The ROC analysis shows that the accuracy of the prediction is 76%.


Patients with schizophrenia require lifelong treatment even when symptoms have improved. Unfortunately, antipsychotic drugs aren't equally effective for every patient; it's unknown who might or might not benefit from the treatment1. Previous studies have found abnormal white matter tracts connecting frontal and temporal structures which were predictive of the short-term outcome of psychosis2. However few have tried to identify the image biomarkers based on the alteration patterns of the whole brain fiber tracts. Therefore, the purpose of this study is to identify the image biomarkers for schizophrenia patients in order to predict treatment responses on individual subject basis.


Ninety-nine patients with schizophrenia were recruited in this study (male/female = 47/52, age = 33.33 ± 9.02). Patients were divided into two groups (remission and non-remission) accordingly to the Positive and Negative Syndrome Scale (PANSS) scores. Diffusion spectrum imaging (DSI) and T1-weighted MRI images were acquired on a 3T MRI system (TIM Trio, Siemens, Germany). T1‐weighted imaging used a MPRAGE sequence (TR/TE = 2000/3 ms, flip angle = 9°, acquisition matrix = 192 × 256, FOV = 256 × 256 mm^2, resolution = 1 x 1 x 1 mm^3). DSI used a pulsed gradient twice‐refocused spin‐echo diffusion echo‐planar imaging sequence (TR/TE = 9600/130 ms, FOV = 200 × 200 mm^2, matrix size = 80 × 80, slice thickness = 2.5 mm)3. A total of 102 diffusion-encoding gradients were applied with the maximum diffusion sensitivity bmax of 4000 s/mm^2. The diffusion-encoding gradients corresponded to the grid points in the diffusion-encoding space (q space) filled in a half sphere4. We used the tract-based automatic analysis (TBAA) to analyze the tract integrity of 76 major fiber tract bundles in the brain5. The output of the TBAA analysis was a 2D connectogram for each DSI dataset, presenting GFA profiles of the 76 tract bundles. To identify the image biomarkers, patients were divided randomly into two groups, a training group and a testing group, with comparable number of remitted and non-remitted patients. The training group was used to compare the difference between remitted and non-remitted patients and to generate treatment response-relevant masks at different thresholds. We used 15 levels of cluster size and 20 levels of effect size as our combined thresholds, resulting in 300 masks. The testing group was used to verify the performance of the prediction for each of the masks using the receiver operating characteristic (ROC) analysis. Specifically, each individual patient’s GFA value was compared to the mean values of the remitted and non-remitted patients at the same step of the connectogram in the training group. If patient’s GFA value was close to remission, the step was set to 0, while it was set to 1 means close to non-remission. Accordingly, we obtained a remission-or-non-remission (RON) map for each subject. The RON map was then masked by the masks obtained from the training group, producing 300 masked RON maps. We devised a scoring system of likelihood of non-remission by counting the total number of non-zero steps in the masked RON map. The number of count was defined as the non-remission index (NRI). Consequently, we obtained the NRI for each of the masks in each subject. Once each subject had the NRI for each mask, we performed the ROC analysis to test the performance of the mask by comparing the NRI and the subject’s final diagnosis (remission or non-remission). We used the area-under-curve (AUC) value to indicate the accuracy of the prediction and repeated the test for200 permutation. The endpoint of the study was to identify the mask that best predicted the non-remission and its accuracy of the prediction.


We found that some masks exhibited high AUC values in predicting remission and non-remission. The masks with the cluster size of 14 to 15 and the effect size of 0.15 ± 0.10 showed the best prediction (Figure 1). By cumulating the best-performance masks over 200 permutations, we obtained the heat map of the segments that contributed significantly to the prediction (Figure 2). These segments were considered as the potential biomarkers of the treatment response in schizophrenia. The mean AUC value of the best-performance masks over 200 permutations was approximately 76%. This meant that using our prediction algorithm on individual subject basis, the accuracy of predicting the treatment response was around 76%.


The algorithm based on the alteration patterns of the whole-brain fiber tracts reveals satisfactory AUC values in predicting the treatment response of schizophrenia (76% accuracy) on individual subject basis.


Future work will involve an independent longitudinal study to validate the clinical value of these biomarkers.


No acknowledgement found.


1. Dazzan, Paola. Neuroimaging biomarkers to predict treatment response in schizophrenia: the end of 30 years of solitude? Dialogues Clin Neurosci. 2014;16(4):491-503.

2. Luck, D. et al. Fronto-temporal disconnectivity and clinical short-term outcome in first episode psychosis: a DTI-tractography study. J Psychiatr Res. 2011;45(3):369-77.

3. Reese, T. G. et al. Reduction of eddy-current-induced distortion in diffusion MRI using a twice-refocused spin echo. Magn Reson Med. 2003;49(1):177-82.

4. Kuo, L. W. et al. Optimization of diffusion spectrum imaging and q-ball imaging on clinical MRI system. Neuroimage. 2008;41(1):7-18.

5. Chen, Y. J. et al. Automatic whole brain tract-based analysis using predefined tracts in a diffusion spectrum imaging template and an accurate registration strategy. Hum Brain Mapp. 2015;36(9):3441-58.


Figure1: The pixels represent the AUC values at different masks. The masks with best prediction are the ones outlined in red.

Figure2: Heat map of the segments with significant contributions to the prediction of treatment response in schizophrenia. The color saturation of the map depends on its occurrence frequency in 200 per permutations. Blue: GFA value of remission is larger than that of non-remission. Orange: GFA value of non-remission is larger than that of remission.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)