Convolutional Sparse Coding Multiple Instance Learning for Whole Slide Image Classification

Multiple Instance Learning (MIL) is commonly utilized in weakly supervised whole slide image (WSI) classification. MIL techniques typically involve a feature embedding step using a pretrained feature extractor, then an aggregator that aggregates the embedded instances into predictions. Current efforts aim to enhance these sections by refining feature embeddings through self-supervised pretraining and modeling correlations between instances. In this paper, we propose a convolutional sparsely coded MIL (CSCMIL) that utilizes convolutional sparse dictionary learning to simultaneously address these two aspects. Sparse dictionary learning consists of filters or kernels that are applied with convolutional operations and utilizes an overly comprehensive dictionary to represent instances as sparse linear combinations of atoms, thereby capturing their similarities. Straightforwardly built into existing MIL frameworks, the suggested CSC module has an affordable computation cost. Experiments on various datasets showed that the suggested CSC module improved performance by 3.85% in AUC and 4.50% in accuracy, equivalent to the SimCLR pretraining (4.21% and 4.98%) significantly of current MIL approaches.