DIGITAL PATHOLOGY & ARTIFICIAL INTELLIGENCE
Increasingly powerful AI algorithms are poised to disrupt all aspects of society, especially healthcare. Data rich clinical specialties, such as pathology, are facing a major paradigm shift. AI will permeate medical diagnostic laboratories in many ways, with analysis of large genetic and transcriptomic dataset being an obvious setting, but one of the less visible yet highly transformational innovations will be the application of AI to digital whole slide images (WSI). Incremental improvements in scanning and data storage (WSIs are orders of magnitude larger than even the radiographic images) have allowed laboratories to incorporate digital pathology into clinical workflows, move away from microscopes and glass slides, and begin the transition to 100% digital workflows just as radiology did beginning in the late 1990s. The implications are profound. There will be tremendous benefits in terms of efficiencies, cost savings, and improvements in quality. Pathologists will be able to view and share WSIs from anywhere in the world, increasing access to pathology expertise in underserved regions. But the most profound shift will arise when AI, propelled by state-of-the-art graphical processing units, is unleashed on WSIs. We are beginning to see the first FDA-authorized AI diagnostic software emerging. Technologies such as these have the potential to automate the most time consuming and subjective aspects of pathology practice, providing more accurate and reliable diagnoses.
The Crary laboratory is on the forefront of using AI to assess the nervous system. We were the first to apply convolutional neural networks to measure Alzheimer’s pathology. We also developed a neural network capable of diagnosing Parkinson’s disease pre-mortem in peripheral biopsies. Yet the potential goes beyond unidimensional feature detection. AIs can “see” structures that are imperceptible or too subtle for the human observer, based on spatial relationships, textures, hues, etc. Further, neural networks can extract essentially limitless numbers of features from histological preparations in an unbiased way and train on them, learning the most relevant structural abnormalities for a given context. When applied to disease, we see enormous potential to unlock the data and knowledge from the huge numbers of unscanned slides that are currently sequestered in archival slide drawers.