Enriching Patient Encoding with Spatial Statistics in Computational Pathology

Deep multiple instance learning is the go-to in computational pathology as it efficiently handles weakly-supervised training. The use of attention mechanism allows to identify the tiles from the whole-slide images (WSIs) that are used to make a prediction, offering a first entry for interpretation. Recent studies proposed unsupervised approaches to summarize WSIs as a mixture of prototypical tiles, giving the opportunity to encode patients as a blend of different morphological patterns that can be translated to biological and clinical interpretation. These methods, albeit promising, fail to capture spatial context between tiles. In this paper, we propose to use spatial statistics to add spatial information to unsupervised patient encodings. More specifically, each tile is associated to its closest prototypical tile and Ripley's K function is used to characterize their spatial homogeneity across the slide. The evaluation of this new patient encoding on the Prostate cANcer graDe Assessment (PANDA) dataset demonstrates its superiority when compared to current patient encodings and MIL methods. Additionally, we show how to interpret this new patient encoding as a specific combination of morphological patterns and their spatial distribution at the WSI level.