Synthetic Cross-sequence Generation of MRI Images using Deep Learning Networks

Abstract

Magnetic Resonance Imaging (MRI) is a non-invasive diagnostic imaging modality that employs multiple acquisition sequences to generate complementary tissue contrasts. However, in clinical practice, not all sequences are available due to time, cost, or patient-related issues. In this work, we investigate deep learning–based supervised image-to-image translation for synthetic cross-sequence MRI generation using a dataset of 148 patients. T2, FLAIR, and contrast-enhanced T1 (T1GD) images are synthesized from T1 inputs using U-Net–based encoder–decoder models and conditional GANs (pix2pix), under full-resolution and patch-based training strategies. Evaluation using Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) shows that UNet models with tailored loss functions achieve results comparable to state-of-the-art approaches, outperforming pix2pix. Finally, qualitative evaluation revealed that conventional metrics fail to capture sequence-specific localized regions, highlighting the need for task-aware evaluation criteria.