Olayinka Oladosu, Rehman Tariq, Shayan Shahrokhi, Y Zhang
Clinical MRI is essential for managing neurological disorders such as multiple sclerosis (MS) but often inconsistent, limiting secondary analyses for enhanced information characterization. Our goal was to establish a new Z-score template method for person-specific image augmentation, as compared to a common deep learning approach termed cycle generative adversarial network (CycleGAN), and test their utility using a treatment response predicting example in MS. We examined 148 MS participants (102 females) from two cohorts, 104 and 44, respectively, with equivalent sex ratio and age, and each included T1-weighted, T2-weighted, and fluid attenuated inversion recovery (FLAIR) brain MRI. The 104/148 participants were used for method development and 44/148 for held-out testing. Z-score templates were constructed using different sample sizes to compare. Z scores from the best template were used to create person- and sequence-specific images. Similar experiments were done using CycleGAN along with tests using the same cohort. Image quality was assessed using peak signal-to-noise ratio, structural similarity index, and root mean square error. Utility testing applied ResNet50-based deep learning models with cycling of the typically unavailable T1-weighted MRI. We found that Z-score template constructed with 75 individuals was the best. Using existing images, potentially unavailable MRI could be created using either method investigated. At an individual level, Z-score template synthesis was equivalent to CycleGAN results. Further, models trained with synthetic or source T1-weighted images achieved similar accuracies (0.82-0.84 ± 0.04 vs. 0.84 ± 0.02) in tests of treatment response prediction. Without using T1-weighted MRI, the model accuracy decreased to 0.72, and AUC decreased below chance. Overall, Z-score template appeared to be a competitive method for person-specific brain MRI augmentation, and synthesized images such as T1-weighted brain MRI have the potential to support downstream applications as seen in prediction of treatment response in MS.