Source-Free Domain Adaptation of Weakly-Supervised Object Localization Models for Histology

TL;DR

This study evaluates source-free unsupervised domain adaptation for weakly supervised object localization in histology, focusing on how domain shifts from staining and scanners affect localization. It compares four white-box SFDA methods—SFDA-Distribution Estimation (SFDA-DE), Source HypOthesis Transfer (SHOT), Cross-Domain Contrastive Learning (CDCL), and AdaDSA—across three WSOL models (DeepMIL, GradCAM++, TS-CAM) using GlaS and Camelyon16 datasets. The findings show classification gains are possible after SFDA, but localization improvements are often limited or inconsistent, highlighting a misalignment between adaptation objectives and WSOL localization. The work emphasizes the critical role of source-model selection (B-LOC vs B-CL) and training strategies, suggesting that future SFDA approaches for histology should jointly optimize localization and classification to enhance interpretability in cancer diagnosis.

Abstract

Given the emergence of deep learning, digital pathology has gained popularity for cancer diagnosis based on histology images. Deep weakly supervised object localization (WSOL) models can be trained to classify histology images according to cancer grade and identify regions of interest (ROIs) for interpretation, using inexpensive global image-class annotations. A WSOL model initially trained on some labeled source image data can be adapted using unlabeled target data in cases of significant domain shifts caused by variations in staining, scanners, and cancer type. In this paper, we focus on source-free (unsupervised) domain adaptation (SFDA), a challenging problem where a pre-trained source model is adapted to a new target domain without using any source domain data for privacy and efficiency reasons. SFDA of WSOL models raises several challenges in histology, most notably because they are not intended to adapt for both classification and localization tasks. In this paper, 4 state-of-the-art SFDA methods, each one representative of a main SFDA family, are compared for WSOL in terms of classification and localization accuracy. They are the SFDA-Distribution Estimation, Source HypOthesis Transfer, Cross-Domain Contrastive Learning, and Adaptively Domain Statistics Alignment. Experimental results on the challenging Glas (smaller, breast cancer) and Camelyon16 (larger, colon cancer) histology datasets indicate that these SFDA methods typically perform poorly for localization after adaptation when optimized for classification.

Figures (21)

• Figure 1: Overall taxonomy of SFDA methods for classification as defined in fang2024source.
• Figure 2: Illustration of SFDA process. (A) The SFDA-DE method is based on distribution estimation. It generates features in the style of the source to perform alignment with target data. (B) CDCL is based on contrastive learning, where positive samples are pulled close while negative ones are pushed apart. In the absence of the source domain, the classifier's weights are used to define prototypes of each class learned on source data. (C) SHOT is a hidden structure mining method based on information maximization. (D) AdaDSA focuses on combining the statistics of the BN layers from the source to normalize target data.
• Figure 3: Visualisation on target (GLAS) dataset with TS-CAM with best localization source model trained with CAMELYON with source's best classification.
• Figure 4: TSCAM best classification on CAMELYON512 without source's best classification.
• Figure 5: TSCAM best classification on GLAS with source's best classification.