Keys to Optimizing AI Model Selection and Integration in Radiology

Integration of artificial intelligence (AI) into a facility’s radiology workflow is an ongoing evolving process. Accordingly, Sriyesh Krishnan, M.D., maintains that one is not only evaluating the potential of a particular AI software or platform, he or she also needs to assess whether there is a good fit with the AI vendor.

“Even if it's just a vendor client relationship, our goal is to find people who are willing to treat it like a partnership because (the AI software is) not going be ready on day one,” noted Dr. Krishnan, the director of clinical AI for Radiology Partners.

During an interview with Diagnostic Imaging at the recent Radiological Society of North America (RSNA) conference, Dr. Krishnan emphasized ongoing evaluation for assessing the adjunctive benefit of AI and to prevent both overreliance and mistrust of the technology.

“We have these checks in place that if someone (is) agreeing too much with the AI or disagreeing too much with AI, we look into why that is happening. Maybe the AI doesn't work well for your site or your patients for some reason. Maybe you're simply distrusting, and we should educate you more on what it's here for and that it's not here to replace you. So these (are) all things we do as part of a comprehensive monitoring program,” explained Dr. Krishnan, a body and emergency radiologist based in Greensboro, N.C. “It's really important because the end is about our patients, and we need to ensure that after we rolled out this AI, our patient care is better, not worse.”

(Editor’s note: For additional coverage of the recent RSNA conference, click here.)

While acknowledging the challenges with AI integration into radiology workflows, Dr. Krishnan said the benefits in patient care are well worth the effort, citing an example of an AI software flagging an obvious head bleed and moving the case to the top of the worklist for review.

“I'm not saying (AI tools are) perfect. There are false positives, there's friction, there's time involved in validating these, but my patients benefit,” maintained Dr. Krishnan. “I think we should embrace that and understand radiologists of tomorrow are going to be supervising AI. It’s our chance to design how we want that to look, or someone else will design it for us.”

(Editor’s note: For related content, see “Can AI Assessment of Non-Calcified Plaque Volume Enhance CT Assessment of MACE Risk Beyond CAC Scoring?,” “Current and Emerging Concepts with LLMs in Radiology: An Interview with Rajesh Bhayana, MD” and “Emerging Directions with Advances in Enterprise Imaging in Radiology.”)

For more insights from Dr. Krishnan, watch the video below.