Challenges of Developing Models for Gigapixel-Scale Pathology

Interpretation of pathology slides is a crucial part of diagnosing disease and an increasingly important part of drug discovery and drug development pipelines. The massive and ever-growing volume of data being produced for each of these areas has led to increased interest in harnessing machine learning in order to produce new biological insights, discover novel biomarkers, steer patient selection for clinical trials, and improve diagnostic accuracy. Effectively training and deploying computer vision models to interpret pathology images must overcome a massive hurdle: pathology slides are typically 100,000 x 100,000 pixels each, many orders of magnitude larger than is typical in computer vision pipelines. However, a number of approaches have been developed over the last few years which have made automatic interpretation of pathology slides not just feasible but a valuable tool used by many healthcare companies around the world. ML Executive; Formerly, VP of AI at PathAI, Butterfly, Nathan Silberman walks attendees through a number of these novel and creative approaches.