P. Vaish (Mei)

Overview

Medical image segmentation models are usually trained on carefully curated datasets, so they often underperform once deployed on messier, real-world clinical data. Standard, visually-consistent augmentation strategies don’t provide enough robustness to handle this kind of distribution shift.

This work systematically studies two augmentation strategies that don’t target any single, specific type of shift: MixUp and Auxiliary Fourier Augmentation (AFA).

Method

MixUp and AFA are integrated into standard nnU-Net training pipelines for cardiac cine MRI and prostate MRI segmentation, and compared against conventional augmentation baselines across a wide range of imaging variations.

Results

Both methods meaningfully improve out-of-distribution generalisation and robustness to imaging variations. They also produce learned feature representations that are more separable and compact, and integrate easily into existing nnU-Net pipelines — making them a practical way to improve reliability in real-world clinical deployment.