Advancing Ulcerative Colitis Monitoring with Deep Learning Models

Project Summary/Abstract The number of practicing pathologists around the world is expected to decrease by as much as 30% over the next two decades, with some of the world’s poorest countries having a ratio of only one pathologist to many hundreds of thousands of people. At the same time, the diagnostic caseload that requires their expertise in clinical trials and hospital settings will continue to grow. The digitization of pathology data, coupled with the use of machine learning techniques for analyzing and scoring the data, provides exciting opportunities to make the field of pathology more efficient and scalable, even as the workforce continues to evolve. Deep learning in particular provides the potential to enhance the interpretation of medical images by improving the detection of image-based biomarkers for a broad range of diseases. Image interpretation plays an important role in patient eligibility and endpoint determination during the course of clinical trials. For patients with ulcerative colitis, the development of trained and reliable algorithms that can help pathologists identify disease progression and response to treatment in a timely and effective manner can provide benefit in two important ways. First, it will help to ensure that the most appropriate score for histological disease severity is being assigned to each image using the Robarts Histopathology Index (RHI) or similar grading scale. Second, it will support a triage process by which images known to contain non- healthy tissues can be prioritized for earlier assessment. Through a unique partnership between Azavea, a geospatial technology and machine learning firm, and Robarts, a clinical trials organization, the proposed research will begin to address these needs by developing deep learning algorithms for histopathology digital image analysis, testing them on machine-readable annotations of medical imagery from previous clinical studies, and exposing them through a metadata- searchable interface that will enable the images to be categorized and quickly accessed by pathologists and others to support reader training and increase communication between multiple readers and sites. In so doing, it will not only help streamline the evaluation of new ulcerative colitis treatments that rely heavily on the image interpretation process, but also provide the foundation for the identification of additional components present in other gastrointestinal disease indications in the future.Project Narrative The proposed research will contribute critical new insights on the reliability, sensitivity, and practicality of machine learning to support gastrointestinal disease detection and evaluation in a clinical trials setting. In pathology, where manual interpretation of images using a microscope has remained relatively unchanged for decades, machine learning provides particular potential to improve the speed and accuracy of diagnoses by reducing the subjectivity that is often inherent in the process.