Robust Cardiac MRI Segmentation under Unseen Vendor Shift via Latent Diffusion-Based Style Augmentation

Abstract

In multi-center cardiac MRI segmentation, models trained under limited acquisition conditions often suffer substantial performance degradation when evaluated on images acquired with unseen scanner vendors. This thesis studies vendor-induced appearance shift using the public M\&Ms OpenDataset, a multi-centre, multi-vendor cardiac MRI benchmark, under a strict vendor-split out-of-distribution (OOD) protocol in which supervised training is restricted to labeled cases from vendors A and B and evaluation is performed on unseen vendors C and D. I first establish a reproducible OOD baseline by fixing preprocessing, an empty-slice policy, and case-level evaluation based on 3D reconstruction from 2D slice predictions. Under this controlled setting, the baseline segmenter achieves strong in-distribution performance on vendors A and B, but exhibits a clear generalization gap on the unseen vendors.To address this gap, I develop a latent diffusion-based style augmentation framework aimed at perturbing image appearance while limiting unintended anatomical drift. Starting from a latent diffusion backbone with cross-attention conditioning, I introduce vendor-conditioned control and parameter-efficient adaptation by injecting LoRA modules into the key and value projections of cross-attention layers. I then connect this controllable generator to supervised segmentation retraining through an offline augmentation pipeline and evaluate whether such generated style variations improve unseen-vendor robustness.Experimental results show that offline diffusion-based augmentation substantially improves in-distribution validation performance on vendors A and B, but does not yield consistent robustness gains on the unseen vendors under static mask reuse. Motivated by this limitation, I introduce a segmentation-aware online augmentation strategy that generates a small set of image-to-image candidates on-the-fly using a frozen, vendor-conditioned diffusion model and selects augmentations by directly evaluating the current segmentation loss. I further analyze a practical degeneracy in which argmin-based selection increasingly prefers the identity input, and mitigate it with a simple random forcing mechanism that maintains non-trivial augmentation usage. Under the strict A/B-only supervision setting, the best online selection-based configuration improves OOD performance on both unseen vendors, increasing mean Dice on vendor D from 0.7801 to 0.7822 and on vendor C from 0.7261 to 0.7339, without any C/D-based checkpoint selection. These findings suggest that coupling augmentation selection to the segmentation objective is more effective than static offline augmentation for improving unseen-vendor robustness under source-only supervision.