DMIA-MIL: Dual-Model Improved Attention Multiple Instance Learning for Classification of Histopathology WSI

Multiple Instance Learning (MIL) is a leading approach for analyzing histopathology whole slide images (WSI), typically divided into key-instance-based and embedding-based methods. While embedding-based MIL often outperforms key-instance-based methods, both face notable challenges: the former may aggregate irrelevant or conflicting features by considering all instances, and the latter risks assigning incorrect pseudo-labels to key instances. Consequently, existing MIL approaches remain imperfect for WSI classification, highlighting the need for improved techniques to enhance accuracy and robustness. To address these limitations, we propose DMIA-MIL, a novel framework that combines key-instance-based and embedding-based strategies. In the key instance identification stage, we introduce an exponential attention mechanism that applies exponential operations to more effectively suppress irrelevant or distracting features within each bag, enabling the model to accurately differentiate positive from negative instances. Subsequently, in the feature embedding stage, we implement a feature distillation approach, which selectively retains the most critical features based on refined attention scores, thereby producing high-quality embedded representations. Empirical results on a well-known WSI dataset demonstrate that DMIA-MIL outperforms existing MIL models and achieves state-of-the-art (SOTA) performance in whole slide image classification. By combining exponential attention and feature distillation, our approach enhances the model's discriminative capabilities and offers a promising direction for tackling the inherent challenges of MIL in large-scale histopathological analysis.