Abstract BACKGROUND Glioma segmentation using deep learning generally relies on additional MRI contrasts that are not required for manual delineation. Retrospective datasets also frequently lack pre-contrast T1-weighted (T1-Pre) or T2-weighted MRI images, or acquire them as a low resolution, limiting the practical deployment of segmentation models. Generating synthetic MRI contrasts through advanced machine learning techniques can overcome this challenge, enabling effective longitudinal monitoring of glioma progression. MATERIAL AND METHODS We developed and validated a conditional Generative Adversarial Network (cGAN) with a distance-based loss function to generate synthetic T1-Pre MRIs from post-contrast images, with an emphasis on enhancing tumor boundaries. The model was trained on 561 and tested on 140 MRIs, with both sets comprising MRIs from the BraTS2020 and institutional datasets. Building on this, we applied transfer learning to fine-tune the pretrained cGAN for synthesizing T2-weighted MRIs from T2-FLAIR sequences, aiming to improve model efficiency and cross-modal generalizability. RESULTS Incorporation of the distance-based loss significantly improved image quality within the lesion, with median Structural Similarity Index Measure (SSIM) and Mean Squared Error (MSE) improving by 10% and 25%, respectively (p 0.00001). The synthetic T1-Pre MRIs achieved median SSIM values of 0.93 for the whole brain and 0.68 for the tumor-only region, with corresponding MSE values of 0.0007 and 0.003. These images also enabled accurate segmentation, yielding median Dice Similarity Coefficients of 0.88 for contrast-enhancing and 0.91 for non-enhancing tumor regions. The fine-tuned model for generating synthetic T2-weighted MRIs demonstrated strong generalization, with a median SSIM of 0.84 for the whole brain and 0.81 for the tumor-only region, and MSE values of 0.007 and 0.004, respectively. CONCLUSION Our cGAN framework, optimized with a tumor-aware distance-based loss function and enhanced through transfer learning, enabled accurate synthesis of missing T1-weighted and T2-weighted MR images. This approach provides consistent lesion segmentations for volumetrics and robust monitoring of glioma progression, particularly in clinical or retrospective scenarios with incomplete MRI exams. SUPPORT This work was supported by the Sandler Program for Breakthrough Biomedical Research, P01CA118816, and T32CA151022 from the National Institutes of Health.
Hadad et al. (Wed,) studied this question.