Data Efficient and Weakly Supervised Computational Pathology on Whole Slide Images

The rapidly emerging field of computational pathology has the potential to\nenable objective diagnosis, therapeutic response prediction and identification\nof new morphological features of clinical relevance. However, deep\nlearning-based computational pathology approaches either require manual\nannotation of gigapixel whole slide images (WSIs) in fully-supervised settings\nor thousands of WSIs with slide-level labels in a weakly-supervised setting.\nMoreover, whole slide level computational pathology methods also suffer from\ndomain adaptation and interpretability issues. These challenges have prevented\nthe broad adaptation of computational pathology for clinical and research\npurposes. Here we present CLAM - Clustering-constrained attention multiple\ninstance learning, an easy-to-use, high-throughput, and interpretable WSI-level\nprocessing and learning method that only requires slide-level labels while\nbeing data efficient, adaptable and capable of handling multi-class subtyping\nproblems. CLAM is a deep-learning-based weakly-supervised method that uses\nattention-based learning to automatically identify sub-regions of high\ndiagnostic value in order to accurately classify the whole slide, while also\nutilizing instance-level clustering over the representative regions identified\nto constrain and refine the feature space. In three separate analyses, we\ndemonstrate the data efficiency and adaptability of CLAM and its superior\nperformance over standard weakly-supervised classification. We demonstrate that\nCLAM models are interpretable and can be used to identify well-known and new\nmorphological features. We further show that models trained using CLAM are\nadaptable to independent test cohorts, cell phone microscopy images, and\nbiopsies. CLAM is a general-purpose and adaptable method that can be used for a\nvariety of different computational pathology tasks in both clinical and\nresearch settings.\n