Abnormality-aware multimodal learning for WSI classification.

Abstract

Whole slide images (WSIs) play a vital role in cancer diagnosis and prognosis. However, their gigapixel resolution, lack of pixel-level annotations, and reliance on unimodal visual data present challenges for accurate and efficient computational analysis. Existing methods typically divide WSIs into thousands of patches, which increases computational demands and makes it challenging to effectively focus on diagnostically relevant regions. Furthermore, these methods frequently rely on feature extractors pretrained on natural images, which are not optimized for pathology tasks, and overlook multimodal data sources such as cellular and textual information that can provide critical insights. To address these limitations, we propose the Abnormality-Aware MultiModal (AAMM) learning framework, which integrates abnormality detection and multimodal feature learning for WSI classification. AAMM incorporates a Gaussian Mixture Variational Autoencoder (GMVAE) to identify and select the most informative patches, reducing computational complexity while retaining critical diagnostic information. It further integrates multimodal features from pathology-specific foundation models, combining patch-level, cell-level, and text-level representations through cross-attention mechanisms. This approach enhances the ability to comprehensively analyze WSIs for cancer diagnosis and subtyping. Extensive experiments on normal-tumor classification and cancer subtyping demonstrate that AAMM achieves superior performance compared to state-of-the-art methods. By combining abnormal detection with multimodal feature integration, our framework offers an efficient and scalable solution for advancing computational pathology.