Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation

TL;DR

This work tackles domain shift in histopathology caused by staining and scanning differences across hospitals. It introduces SCDA, a domain adaptation framework that combines supervised contrastive learning with a cross-domain constraint within a MIL-based slide representation, coupled with MI-SimpleShot for prototype-based classification. The method uses patch-level PLIP features and batch-global pooling to obtain slide representations, which are aligned across centers and then classified with class prototypes. Experiments on a two-center skin cancer WSI dataset (608 slides, six subtypes) show that SCDA outperforms staining normalization and non-adapted baselines, including in few-shot settings, indicating strong cross-center generalization and potential for efficient deployment in multi-center clinical workflows.

Abstract

Domain shift in the field of histopathological imaging is a common phenomenon due to the intra- and inter-hospital variability of staining and digitization protocols. The implementation of robust models, capable of creating generalized domains, represents a need to be solved. In this work, a new domain adaptation method to deal with the variability between histopathological images from multiple centers is presented. In particular, our method adds a training constraint to the supervised contrastive learning approach to achieve domain adaptation and improve inter-class separability. Experiments performed on domain adaptation and classification of whole-slide images of six skin cancer subtypes from two centers demonstrate the method's usefulness. The results reflect superior performance compared to not using domain adaptation after feature extraction or staining normalization.

Figures (3)

• Figure 1: Method overview. In this article, we address inter-center domain shifts through supervised contrastive learning. Concretely, a constraint is introduced in the model training by forcing samples of the same class to be closely spaced. The inclusion of this condition strengthens inter-class clustering and removes inter-center variability.
• Figure 2: In-depth study of SCDA under the few-shot learning paradigm. In (a) HUSC train set is used, and HCUV samples are added. In (b), the reverse procedure is evaluated. The results of the HUSC test set are shown in blue and those of HCUV in orange. The starting point on the curves refers to the 0-shot case, where only the PLIP and BGAP feature extractors are used to predict. The final triangle in the curves refers to the case where both databases are trained. The orange horizontal dotted line indicates the results for the HUCV test set trained with the two databases without using SCDA. The purple dotted line refers to the same but with the HUSC dataset.
• Figure 3: 2D t-SNE representations of feature embeddings extracted by PLIP and aggregation via BGAP. The subfigures illustrate (a) Representation for the embeddings without any added processing. (b) Representation of the embeddings obtained with PLIP and BGAP by previously applying the Macenko normalization method. (c) Representation of the embeddings by applying SCDA.