Exploration of Deep Learning Methods for Synthetic T2-Weighted Pelvic MRI Generation from CT Scans: A Technical...

Synthesizing T2-weighted MRI from CT scans presents a challenging ill-posed problem that remains underexplored in abdominopelvic imaging. We aim to develop and compare deep learning algorithms for generating synthetic T2-weighted MRI from pelvic CT, systematically evaluating architecture and training strategies for feasibility and performance. The framework adopts a conditional Generative Adversarial Network (GAN) approach. Three state-of-the-art models [efficient Self-Attention UNet (ESAUNet), Residual Vision Transformer (ResViT), and Cascaded Gaze] were utilized as generators. A combined loss function (L1, VGG19 perceptual, adversarial) was employed to optimize fidelity. A multicenter cohort (n = 90) including the SynthRad2023 (n = 60) and in-house (n = 30) datasets was used for model training. Five-fold cross-validation was used, and independent testing was performed on the Gold Atlas male cohort (n = 19). Blinded qualitative assessment by two radiologists was conducted. Rigorous qualitative and quantitative assessments were performed. On the held-out cohort, ESAUNet achieved the highest average performance across all metrics: PSNR 22.21 dB, SSIM 0.748, and MAE 0.044. This surpassed the performance of the ResViT model (PSNR 21.18 dB, SSIM 0.727) and CascadedGazeNet (PSNR 21.41 dB, SSIM 0.703). Radiologists found no significant difference in overall pelvic image quality (p = 0.294) or rectal wall delineation (p = 0.358) between synthetic T2-weighted MRI and true MRI, despite lower ratings for fine details. Inter- and intra-observer agreement was robust. Conditional GAN architectures demonstrate technical feasibility for CT-to-T2-weighted pelvic MRI translation. ESAUNet's efficiency and robust performance highlight its potential as a key enabler for MRI-equivalent imaging in resource-limited settings.