SD-MIL: Multiple instance learning with dual perception of scale and distance information fusion for whole slide image classification

Abstract

In computer-aided pathology diagnosis, multiple instance learning (MIL) has become a key method for addressing disease diagnosis problems in whole slide images (WSIs). However, current MIL models have limitations in capturing dependencies among instances and local contextual information. Additionally, the imbalance in the number of positive and negative instances affects MIL models’ ability to identify important instances. To address these issues, we propose a dual perception of scale and distance information fusion method (SD-MIL). SD-MIL consists of two parts: multi-scale window regional self-attention (MWRSA) and adaptive prototype distance-guided instance feature enhancement (PGFE). MWRSA utilizes three different-sized windows to compute regional multi-head self-attention (R-MSA) obtaining scale-aware instance features. This part explores instance long-range dependencies in local region and capture local contextual information at different scales. In the PGFE part, the distance parameter between instances and bag-level prototype is considered to assign different significance weights to instances resulting in distance-aware instance features, which guides model better focus on important instances. Then, learnable parameters optimize the fusion of scale-aware and distance-aware instance features, enhancing instance feature representation and ensuring the downstream aggregation model to generate high-quality bag features. Experimental results on three datasets show that SD-MIL outperforms state-of-the-art MIL methods. Meanwhile, SD-MIL consistently delivers performance improvements when the feature extraction network or downstream aggregation model is replaced.