Abstract
Incomplete MRI exams reduce the possibility of applying AI models. We previously trained two generative adversarial networks (GANs) on 135 subjects from the BraTS-2017 dataset to generate T1 from post-contrast T1 and FLAIR from T2 sequences (baseline models). To test the impact of the training set size, here, we trained the same models using 1251 subjects from the BraTS-2021 dataset. We used the same architecture and hyperparameters from the baseline models: batch size of 1, learning rate 2x10(-4), Adam optimizer, cross-entropy and L1-loss. We trained two versions of the updated models: one stopped at an early checkpoint (early models) and one after 50 epochs (late models). We tested all models on institutional MRIs from 487 IDH-wt gliomas. Synthetic MRIs were compared to the originals using the structural-similarity-index (SSI) and MSE. We simulated scenarios where either the T1, FLAIR, or both were missing and used their synthetic version as inputs of a segmentation model. We compared the segmentations using the dice score (DSC) and used Friedman and Dunn’s test to compare the scores and correct for multiple comparisons. Median SSI for the T1 were 0.957, 0.947, 0.947, and median MSE were 0.006, 0.014, 0.008 for the baseline, early, and late models. For the FLAIR, median SSI were 0.924, 0.908, 0.915, and median MSE were 0.016, 0.016, 0.019 for the baseline, early, and late models. The DSC ranges for the baseline and updated models were (0.655 - 0.953), (0.420 - 0.952), and (0.610 - 0.952). The baseline and late models showed no statistically significant differences in DSC, but both performed significantly better than the early models. We show that GANs trained to synthesize brain MRIs on a small cohort perform similarly to those trained on an x10-larger cohort, making them a viable option for rare diseases or institutions with limited resources.