Advancing Cross-Organ Domain Generalization with Test-Time Style Transfer and Diversity Enhancement

TL;DR

This work tackles domain shifts in computational pathology by introducing Test-time Style Transfer (T3s), a foundation-model-based framework that bidirectionally projects source and unseen domains into a shared style space. A Cross-domain Style Diversification Module (CSDM) expands the style representation space while enforcing orthogonality among style bases, and data augmentation plus low-rank adaptation improve feature alignment across multi-domain WSIs. Empirical results on unseen multi-organ datasets show state-of-the-art performance (e.g., IoU $77.68\%$ and Dice $87.43\%$) compared to baselines such as TT-DG and DCAC, with ablations confirming the contribution of each component. The approach advances cross-organ generalization in pathology segmentation, enabling more robust deployment across diverse organ tissues and institutions.

Abstract

Deep learning has made significant progress in addressing challenges in various fields including computational pathology (CPath). However, due to the complexity of the domain shift problem, the performance of existing models will degrade, especially when it comes to multi-domain or cross-domain tasks. In this paper, we propose a Test-time style transfer (T3s) that uses a bidirectional mapping mechanism to project the features of the source and target domains into a unified feature space, enhancing the generalization ability of the model. To further increase the style expression space, we introduce a Cross-domain style diversification module (CSDM) to ensure the orthogonality between style bases. In addition, data augmentation and low-rank adaptation techniques are used to improve feature alignment and sensitivity, enabling the model to adapt to multi-domain inputs effectively. Our method has demonstrated effectiveness on three unseen datasets.

Figures (3)

• Figure 1: Left: Previous DG method, which learns domain-invariant features during training; Right: Test-time Style Transfer method, which projects the source domain and unseen domain into a style representation space, thereby enhancing model generalization on Cross-organ tasks.
• Figure 2: An overview of Test-Time Style Transfer (T3s) framework for Cross-organ tasks. First, we leverage pre-trained foundation models to extract robust, domain-independent features across domains. During training, we introduce Test-Time Style Transfer, a bidirectional mapping mechanism that maps the source and target domains into a style representation space. We then design a Cross-domain Style diversification module (CSDM) to maximize the style representation space. At test time, a domain adaptation strategy dynamically adjusts the style of the test data to align it with the learned features, thereby achieving effective generalization to unknown domains.
• Figure 3: Features of different domains are visualized with T-SNE before (a) and after (b) test-time style projection. Training domain: $\bullet$ Pancreas, $\bullet$ Colorectum, $\bullet$ Stomach; Testing domain: $\star$ Pancreas, $\star$ Colorectum, $\star$ Stomach, $\star$ Intestine, $\star$ Ampullary, $\star$ Gallbladder