SparseXMIL: Leveraging sparse convolutions for context-aware and memory-efficient classification of whole slide images in digital pathology.

The computer analysis of Whole Slide Images (WSIs) has become increasingly prevalent in pathology-based diagnosis, although their analysis still presents considerable challenges due to the voluminous nature of the data. To address this issue, Multiple Instance Learning (MIL) has emerged as a viable approach in which WSIs are partitioned into tiles for processing. Nevertheless, previous MIL methodologies inadequately capture the essential spatial context between tiles, which is imperative for accurate diagnosis across various diseases. In parallel, approaches based on adapting Convolutional Neural Network (CNN) architectures to the high-resolution setting have also emerged. However, these methods are generally GPU memory-intensive, restricting the architecture's size and scalability. In this paper, we present a novel framework, SparseXceptionMIL (SparseXMIL), aiming to enhance the GPU efficiency of spatial interaction modeling within WSI data through the introduction of a multidimensional sparse image representation and a novel pooling operator. By integrating sparse convolutions within the Xception architecture, this operator enables efficient spatial information modeling at both local and global scales. Empirical evaluations conducted on various classification tasks, encompassing subtyping for breast and lung carcinomas and predicting abnormalities in the DNA damage response in breast cancer WSIs, demonstrate that our approach outperforms state-of-the-art MIL methods in tasks where spatial context is important, and offers a better trade-off in terms of GPU memory requirements compared to CNN-based methods. These results underscore the potential of sparse convolutional architectures for efficient and scalable WSI analysis. Our experiments' source code is available at https://github.com/loic-lb/SparseXMIL.