Collaborative optimization for whole slide image classification

Effective instance representation and bag prediction are critical in multi-instance learning (MIL) for histopathology whole slide image (WSI) analysis. Most current methods focus on elaborating either the instance aggregators or bag predictors, while neglecting the synergistic promotion between these aspects during optimization. To mitigate this gap, we proposed COMIL, a unified framework that jointly enhances instance representation and bag prediction through collaborative learning. COMIL employs an instance fuser to capture correlations among instances, a neighbor contrastive learning module to further derive reliable and generalized representations, and a transformer-based aggregator to integrate high-quality features for robust bag prediction. By jointly optimizing these components within a multi-task learning paradigm, COMIL effectively explores and leverages the mutual information between representation and prediction, thus improving overall optimization quality. Extensive experiments on four benchmark datasets demonstrate that COMIL outperforms state-of-the-art methods, achieving accuracy improvements of 1.2 % and 3.3 % over the pseudo-bag-based DTFD, and 4.5 % and 2.0 % over the bag-based ACMIL, on the CAMELYON16 and TCGA-ESCA datasets, respectively, highlighting the effectiveness of COMIL in WSI classification tasks.