Ep. 37 Practical AI: Learning the Basics with Dr. Amit Gupta

In this episode, Dr. Bryan Hartley interviews cardiothoracic radiologist Dr. Amit Gupta about his involvement in artificial intelligence (AI), what it takes to build and implement an AI algorithm, and how radiologists can empower themselves to lead the clinical aspects of this revolution. --- EARN CME Reflect on how this Podcast applies to your day-to-day and earn AMA PRA Category 1 CMEs: https://earnc.me/kuYlpg --- SHOW NOTES Dr. Gupta first became involved in AI as a neuroradiology fellow, mainly with annotating images. Overtime, he fostered good relations with a biomedical engineering team at Case Western. He found that he was able to provide valuable clinical insights for their algorithm development. Dr. Gupta had always loved testing out new technologies, but he realized that he could actually participate in taking science forward and developing artificial intelligence. We review the basics of AI, starting by defining machine learning, deep learning, and convolutional neural networks. Dr. Gupta speaks about all the steps embedded in an ideal algorithm – it would segment a lesion, integrate clinical information, track past growth, evaluate response criteria, and then ultimately tell you likelihood that a patient will respond to a certain therapy. We also cover radiomics, a subset of AI image analysis that looks beyond what humans can see with naked eye and examines the microenvironment of the image. Dr. Gupta outlines the process of AI algorithm development, which includes definition of a clinical problem, data selection and preparation, annotation, testing, and finally, implementation. Since it is now more and more common for radiologists to be approached by AI vendors, we discuss important considerations when working with them. Dr. Gupta emphasizes that from the very beginning of the vendor relationship, radiologists need to evaluate whether or not the proposed technology would solve a clinically relevant problem and improve radiology workflow. If not, testing the algorithm may not be worth the time or effort. Just because an algorithm has FDA approval does not mean that it is clinically useful for a certain institution or geographic population of patients and radiologists. Next, the vendor and radiologist must come to an agreement on deliverables and compensation for testing team members. This could take the form of monetary compensation or an in-kind transfer of the software after it is built. Throughout the testing process, it is important to involve the institution’s PACS, IT, and security teams to ensure images are properly de-identified and patients are protected. Finally, we look at the potential impact that AI has on the diagnostic radiology workflow. Dr. Gupta believes that AI could be rebranded as “assistive technology,” since it would help, rather than replace radiologists. Despite the remarkable capabilities of this technology, it still requires humans to integrate clinical and prior imaging information to make sense of radiographic findings. At Case Western, he fosters a supportive environment for trainees to learn about AI and collaborate with teams that are building the algorithms. --- RESOURCES RSNA Artificial Intelligence Journal: ​​https://pubs.rsna.org/journal/ai ACR Data Science Institute: https://www.acrdsi.org/ Kaggle: https://www.kaggle.com/