Deep learning from multiple experts improves identification of amyloid neuropathologies

Abstract

AbstractPathologists can have complementary assessments and focus areas when identifying and labeling neuropathologies. A standardized approach would ideally draw on the expertise of the entire cohort. We present a deep learning (DL) framework that consistently labels cored, diffuse, and cerebral amyloid angiopathy (CAA) neuropathologies using expert consensus. We collected 100,495 annotations, comprising 20,099 candidate neuropathologies from three institutions, independently annotated by five experts. We compared DL methods that learned the annotation behaviors of individual experts (AUPRC=0.67±0.06 cored; 0.48±0.06 CAA) versus those that reproduced expert consensus, yielding 8.9-13% improvements (AUPRC=0.73±0.03 cored; 0.54±0.06 CAA). Saliency mapping on neuropathologies illustrated how human expertise may progress from novice to expert. In blind prospective tests of 52,555 subsequently expert-annotated images, the models accurately labeled pathologies similar to their human counterparts (consensus model AUPRC=0.73 cored; 0.68 CAA).