AdverIN: Monotonic Adversarial Intensity Attack for Domain Generalization in Medical Image Segmentation

Zheyuan Zhang, Ulas Bagci

Abstract: Domain generalization (DG) has emerged as a promising research direction because it can potentially enable deep learning models to handle data from previously unseen domains. DG methods try to achieve this by learning domain-invariant features that are robust to variations across different domains. This work proposes a novel domain generalization (DG) technique termed Adversarial Intensity Attack (AdverIN). AdverIN leverages an adversarial training strategy to augment data diversity by synthesizing a spectrum of intensity variations while preserving essential contextual information within images. We evaluated AdverIN through rigorous experiments on diverse multi-domain tasks: 2D retinal optic disc/cup segmentation and 3D prostate MRI segmentation. Our results demonstrate that AdverIN significantly improves the generalizability of segmentation models, achieving state-of-the-art performance on these challenging datasets. The code is available at https://github.com/NUBagciLab/AdverIN