GRASP: GRAph-Structured Pyramidal Whole Slide Image Representation

• DOI: 10.48550/arxiv.2402.03592
• Publication Date: 2024-02-05
• Authors: Ali Khajegili Mir...; G. C. Archibald; Amirali Darbandsari; Alberto Contreras...; Ramin Ebrahim Nakhli; Maryam Asadi; Allen Zhang; C. Blake Gilks; Peter McL. Black; Gang Wang; Hossein Farahani; Ali Bashashati
• Keywords: Representation

Cancer subtyping is one of the most challenging tasks in digital pathology, where Multiple Instance Learning (MIL) by processing gigapixel whole slide images (WSIs) has been spotlight recent research. However, MIL approaches do not take advantage inter- and intra-magnification information contained WSIs. In this work, we present GRASP, a novel graph-structured multi-magnification framework for WSIs pathology. Our approach designed to dynamically emulate pathologist's behavior handling benefits from hierarchical structure which introduces convergence-based node aggregation instead traditional pooling mechanisms, outperforms state-of-the-art methods over two distinct cancer datasets margin up 10% balanced accuracy, while being 7 times smaller than closest-performing model terms number parameters. results show that GRASP dynamic finding consulting with different magnifications cancers reliable stable across hyperparameters. The model's evaluated expert pathologists confirming interpretability dynamic. We also provide theoretical foundation, along empirical evidence, our explaining how interacts nodes graph make predictions. believe strong characteristics yet simple will encourage development interpretable, structure-based designs WSI representation Furthermore, publish large rare Ovarian Bladder contribute field.