SCH: Intelligent Radiology Through Human-Machine Cooperation

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

With unprecedented progress in the development of advanced imaging tools for medicine, accurate interpretation of medical images has become essential in diagnosis and treatment of several diseases in a wide range of medical disciplines. Nevertheless, studies show large variability between interpretations of different radiologists, especially using newer developed quantitative scales. Recent computational and algorithmic advances in Artificial Intelligence (AI) promise effective tools to learn the relationship between medical images and clinical data to diagnose diseases and predict progression and outcomes based on the past history of each patient. However, the AI tools for medical analysis are usually developed based on radiologist interpretations, not considering the between-person variations. Vice versa, there is very limited data on how AI can affect radiologists’ readings and offer additional value to their daily workflow. This project is focused on the creation of a trustworthy reference standard for AI using the interpretation of multiple radiologists while investigating how AI can reduce the variability in interpretation and improve the clinical workflow to optimally benefit patient care. It will considerably reduce the variations among radiologist readings by proving the AI assessment while improving the AI algorithms using radiologists’ reading strategy. It will be helpful in training radiologists as well as mitigating the adverse effects of physical, psychological, and environmental conditions (e.g., noise, fatigue) on the radiologists’ readings of medical images. The overarching goal of this project is to develop a novel paradigm based on human-AI cooperation in intelligent computational analysis of medical imaging data by using state-of-the-art AI algorithms to develop an accurate labeling tool for medical images to maximize the accuracy and minimize the inter- and intra-reader variability. The tool also provides the decision-making rationale in the AI algorithm through a series of filtered images to help radiologists in their interpretations. In parallel, an eye-tracking system is used to learn the decision-making patterns of expert radiologists and trainees with and without knowing the AI assessment. This knowledge is fed back to the AI algorithm to improve its performance. This will result in the most reliable labeling tool with superior performance compared to conventional AI tools or individual radiologists. Using this platform, an AI tool is developed for combining the medical imaging data with all relevant clinical, social, and demographic data for accurate diagnosis and prediction of the course of a disease with and without treatment for each patient. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.