SparseConvMIL: Sparse Convolutional Context-Aware Multiple Instance Learning for Whole Slide Image Classification

Multiple instance learning (MIL) is the preferred approach for whole slide\nimage classification. However, most MIL approaches do not exploit the\ninterdependencies of tiles extracted from a whole slide image, which could\nprovide valuable cues for classification. This paper presents a novel MIL\napproach that exploits the spatial relationship of tiles for classifying whole\nslide images. To do so, a sparse map is built from tiles embeddings, and is\nthen classified by a sparse-input CNN. It obtained state-of-the-art performance\nover popular MIL approaches on the classification of cancer subtype involving\n10000 whole slide images. Our results suggest that the proposed approach might\n(i) improve the representation learning of instances and (ii) exploit the\ncontext of instance embeddings to enhance the classification performance. The\ncode of this work is open-source at {github censored for review}.\n