Systems, methods, and apparatuses for learning foundation models from anatomy in medical imaging for use with medical image classification and segmentation

Abstract

Human anatomical structures are extracted from medical images. A foundation model is generated by learning the human anatomical structures, resulting in generic representations of the human anatomical structures. The human anatomical structures are learned by executing a self-supervised contrastive learning framework to conserve hierarchical relationships of the human anatomical structures in the medical images and to capture distinct representations for different human anatomical structures at different granularity levels. The self-supervised contrastive learning framework executes an anatomy decomposer (AD) to conserve hierarchical relationships of the human anatomical structures in the received medical images and executes a purposive pruner (PP) to capture distinct representations for different human anatomical structures at different granularity levels. Medical images that form no part of the received medical images may then be processed using the trained foundation model.