| MS Seminar

Name of the Speaker: Ms. Sadhana S (EE22S041)
Guide: Dr. Mohanasankar S
Online meeting link: http://meet.google.com/aoz-pdii-odo
Date/Time: 9th May 2025(Friday), 2:30 PM
Title: DCEtriformer: Enhancing Prostate DCE-MRI Synthesis with Hybrid Attention Mechanisms.

Abstract :

Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) plays a critical role in prostate cancer diagnosis, offering high sensitivity in lesion characterization, biopsy guidance, and tumor grading. However, DCE-MRI requires contrast agents, which may not always be feasible due to cost, time, or patient risk. This thesis explores the synthesis of early- and late-response DCE-MRI images using non-contrast multiparametric MRI (mpMRI) modalities, including T2-weighted (T2W), Apparent Diffusion Coefficient (ADC), and pre-contrast DCE images.

We first present DCE-former, a ransformer-based generative adversarial network (GAN) built on a U-Net-like architecture. It is trained using (i) 1 loss, (ii) mutual information loss to yield high semantic similarity, and (iii) frequency-based loss to reserve spatial and spectral fidelity in contrast-enhanced regions at both local and global levels.

To enhance feature diversity, we propose DCEtriformer, an extension of DCE-Former that alternates triangular and rectangular attention within layers based on LeWin block depth. Variable window sizes are also employed to jointly learn local and global contrast dynamics. This hybrid attention approach improves spatial coverage and enhances robustness to domain shifts.

Experiments on PROSTATEx and Prostate-MRI datasets show that DCEtriformer surpasses existing methods in synthesizing linically relevant DCE-MRI images, demonstrating the potential of deep transformer-based models for contrast-free imaging in prostate cancer.