Refining Multiple Instance Learning with Attention Regularization for Whole Slide Image Classification
Abstract: Histopathological analysis of biopsy sections is crucial for the detection of cancer and the distinction between different tumor subtypes. To this end, pathologists identify certain key regions of the biopsy from which a diagnosis is derived. The classification of whole slide images (WSIs) can be addressed as a multiple instance learning (MIL) problem where only slide-level labels are available. In order to model the relevance scores of the different WSI patches, attention mechanisms are implemented within the MIL framework. However, excessive flexibility in the attention mechanisms for computing attention scores to patches may lead to nearly uniform attention distributions, potentially deteriorating the model’s performance. In this paper, we introduce an unsupervised auxiliary loss function to recalibrate the attention mechanism enhancing emphasis on crucial patches and downscaling the influence of less relevant ones. The proposed MIL framework has been evaluated on invasive breast carcinoma (TCGA-BRCA) and renal cell carcinoma (TCGA-RCC) subtyping. The results obtained show that attention regularization not only improves the predictive capacity of the model but also significantly increases its interpretability by identifying regions and patterns of high diagnostic value.
Loading