Abstract: Deep learning segmentation models degrade when applied across imaging modalities like MRI and CT, where differences in acquisition physics lead to stark contrasts in intensity and boundary appearance. Unsupervised domain adaptation (UDA) addresses this without requiring labeled target data, but existing prototype-based approaches suffer from unstable class anchors that drift as features and pseudo-labels evolve during training. This thesis proposes Equiangular Prototype Alignment (EPA), a feature-level UDA objective that anchors class representations to a fixed set of simplex Equiangular Tight Frame (ETF) prototypes -geometrically balanced, unit-norm vectors at equal pairwise angles that remain frozen throughout training. EPA is integrated with mean-teacher self-training and Fourier Domain Adaptation (FDA) to simultaneously address domain shift at the supervision, appearance, and representation levels.

Evaluated on the MM-WHS cardiac benchmark across four structures, EPA outperforms the strongest prior UDA baseline in both adaptation directions reaching 85.2% mean Dice and 3.3 mean ASD for MR→CT (+1.1 Dice, +0.5 ASD, with a 7.5-point gain on the difficult myocardium) and 71.2% mean Dice and 2.2 mean ASD for CT→MR (+1.3 Dice, +1.6 ASD). Ablation confirms that FDA provides the largest single gain by closing the low-level appearance gap, while EPA contributes 14.2 Dice points without FDA and 7.3 with it. Together, these results show that fixed ETF prototypes offer a stable, drift-free geometric reference that converges faster than dynamic prototype variants making explicit representation geometry a practical, low-overhead strategy for cross-modality medical image segmentation.

Event Details
Title: Equiangular Prototype Alignment: Fixed Simplex-ETF Geometry for Unsupervised Domain-Adaptive Medical Image Segmentation
Date: June 23, 2026 at 02:30 PM
Venue: Google Meet (https://meet.google.com/jic-dbqw-cpw)
Speaker: Mr. Akash Sharma (EE21S056)
Guide: Dr. Mohanasankar S
Type: MS seminar

Updated: