Title: AI Innovation in Medical Imaging: Where Are We?
Host: Department of Biomedical Informatics, University of Pittsburgh
Speaker: Shandong Wu
Summary and thoughts:
Dr. Shandong Wu from the Department of Biomedical Informatics (DBMI) and Pittsburgh Center for AI Innovation in Medical Imaging presented at this week’s colloquium. His talk was divided into two major parts - recent innovations in medical imaging in general and then focused on particular projects in his lab. I don’t have any experience working with medical images but it is an interesting area adjacent to my current work with electronic health records. Imaging forms 40% of all healthcare procedures including pathology, radiology and ophthamology. Unsurprisingly, with advances in artificial intelligence (AI) and computer vision, there has been a major boost in AI imaging technology in the past 5 years. Centers all over the country, both in government organizations and universities (including University of Pittsburgh) are dedicated to medical imaging research. Dr. Wu leads the Pittsburgh Center for AI Innovation in Medical Imaging (https://www.aimi.pitt.edu/), a collaboration between Pitt, CMU and UPMC with over 90 team members. The center aims to advance innovation and research in medical imaging as well as work on the many challenges in the field. These include challenges with data (limited labels, heterogeneity, security, sharing protocols), algorithms (bias, reproducibility, explainability, deployment), and the workforce (academic-industry synergy, clinical translation, multi-disciplincary research). Dr. Wu anticipates the center to be heavily involved in interdisciplinary research through conversing fields, with particular inputs from clinical experts.
His research lab currently focuses on solving problems in the AI spectrum in a way that makes the systems trustworthy for the users. This is done by involving clinical knowledge in all algorithms such as deep learning methods. The results of several projects such as breast cancer risk assessment, patient stratification, survival analysis, head CT image interpretation and outcome predication after cardiac arrest, and more have shown that algorithms perform much better when integrated with domain knowledge such as clinical insight. Some other ways the lab is overcoming challenges include data harmonization, identification of model safety from adversial attacks, external validation and multi-modal machine learning. While the innovations and methods for medical imaging continue to improve, there is a long way to go before trustworthy AI can be fully deployed in the healthcare system. I believe there is an even longer way to go before it can adapt to the other systems (such as insurance, legal, administration) closely tied to the healthcare system in the country.