Bridging the Inter-Domain Gap through Low-Level Features for Cross-Modal Medical Image Segmentation

TL;DR

This work tackles cross-modal medical image segmentation under unsupervised domain adaptation by proposing LowBridge, a model-agnostic framework that leverages edge features as domain-invariant cues. A simple edge-to-image generator reconstructs source-style images from edge maps, enabling a segmentation model trained on the source domain to operate on the target domain without test-time fine-tuning. Key contributions include a straightforward training scheme, state-of-the-art results on CHAOS (liver) and MMWHS (cardiac) benchmarks, and extensive ablations demonstrating robustness across different generator and segmentation backbones. The approach highlights the potential of low-level feature invariances to bridge modality gaps and paves the way for combining with more advanced generative techniques in future work.

Abstract

This paper addresses the task of cross-modal medical image segmentation by exploring unsupervised domain adaptation (UDA) approaches. We propose a model-agnostic UDA framework, LowBridge, which builds on a simple observation that cross-modal images share some similar low-level features (e.g., edges) as they are depicting the same structures. Specifically, we first train a generative model to recover the source images from their edge features, followed by training a segmentation model on the generated source images, separately. At test time, edge features from the target images are input to the pretrained generative model to generate source-style target domain images, which are then segmented using the pretrained segmentation network. Despite its simplicity, extensive experiments on various publicly available datasets demonstrate that \proposed achieves state-of-the-art performance, outperforming eleven existing UDA approaches under different settings. Notably, further ablation studies show that \proposed is agnostic to different types of generative and segmentation models, suggesting its potential to be seamlessly plugged with the most advanced models to achieve even more outstanding results in the future. The code is available at https://github.com/JoshuaLPF/LowBridge.

Figures (2)

• Figure 1: Framework of our LowBridge framework. Edge features are treated as domain-invariant representations to train a generative model, $\mathcal{G}$, to generate source-style images. Edge features extracted from unlabeled target images are then input to $\mathcal{G}$, followed by a segmentation model, $\mathcal{S}$, pretrained on the source data, to output the predictions.
• Figure 2: Visualization results on the CHAOS and MMWHS datasets.