Dual attention model with reinforcement learning for classification of histology whole-slide images

Digital whole slide images (WSIs) are generally captured at microscopic resolution and encompass extensive spatial data (several billions of pixels per image). Directly feeding these images to deep learning models is computationally intractable due to memory constraints, while downsampling the WSIs risks incurring information loss. Alternatively, splitting the WSIs into smaller patches (or tiles) may result in a loss of important contextual information. In this paper, we propose a novel dual attention approach, consisting of two main components, both inspired by the visual examination process of a pathologist: The first soft attention model processes a low magnification view of the WSI to identify relevant regions of interest (ROIs), followed by a custom sampling method to extract diverse and spatially distinct image tiles from the selected ROIs. The second component, the hard attention classification model further extracts a sequence of multi-resolution glimpses from each tile for classification. Since hard attention is non-differentiable, we train this component using reinforcement learning to predict the location of the glimpses. This approach allows the model to focus on essential regions instead of processing the entire tile, thereby aligning with a pathologist's way of diagnosis. The two components are trained in an end-to-end fashion using a joint loss function to demonstrate the efficacy of the model. The proposed model was evaluated on two WSI-level classification problems: Human epidermal growth factor receptor 2 (HER2) scoring on breast cancer histology images and prediction of Intact/Loss status of two Mismatch Repair (MMR) biomarkers from colorectal cancer histology images. We show that the proposed model achieves performance better than or comparable to the state-of-the-art methods while processing less than 10% of the WSI at the highest magnification and reducing the time required to infer the WSI-level label by more than 75%. The code is available at github.