Suggestions
Woojin Kim
Chief Medical Information Officer at Rad AI | CMO at ACR DSI | MSK Radiologist | Imaging Informaticist | Entrepreneur | Consultant | Speaker | AI Artist | Travel Photographer
Woojin Kim is a seasoned professional in the healthcare industry, currently serving as the Chief Medical Information Officer (CMIO) at Rad AI. With a background spanning over 16 years as a diagnostic radiologist, Woojin specializes in musculoskeletal radiology and imaging informatics. His role involves developing AI-powered solutions to streamline radiology processes, enhance patient care, and combat burnout among radiologists. Beyond his medical expertise, Woojin is an accomplished entrepreneur with experience in founding and successfully exiting three healthcare startups focused on data mining, clinical analytics, machine learning, and more.
Highlights
🎨 Radiological Society of North America (RSNA) brought back a radiology-themed art contest with three categories: Radiology Art, Connections Through Art, and Artificial Intelligence Art. Top picks will be displayed at #RSNA24.
🗳️ Voting has started!
👨🎨 I participated in the AI Art category this year. 👇
😊 Visit the #RSNAContest webpage, view all the submitted artwork, and cast your vote 👉 https://buff.ly/3MmOlcj
#radiology #RadiologyArt #AIart
🤔 Where are we with automated radiology reporting?
📝 An interesting paper was published in Radiological Society of North America (RSNA) Radiology by researchers in Denmark using AI to identify "unremarkable" chest X-rays (CXRs) for automatic reporting. When the AI model from Annalise.ai (used off-label) operated at 87.2% sensitivity, similar to the radiology reports, it produced more critical misses than the radiology reports (2.2% vs. 1.1%). However, when the AI model operated at a sensitivity of 95.4% or higher, it achieved an equal or lower critical missed rate (≤1.1%). At this latter sensitivity level, they would have seen a case volume reduction of 23.5%.
💡 While several limitations exist, including the need to demonstrate this prospectively, this isn't the first study to show the potential of automated reporting in radiology using a commercial solution and will not be the last.
🤖 In research, models like Med-PaLM M, ELIXR, RayDINO, RAD-DINO, and MAIRA-2 have shown automated reporting of CXRs beyond filtering for normal/unremarkable exams. At RSNA 2023, Rad AI demonstrated a proof-of-concept automated CXR reporting beyond normal/unremarkable exams using Google’s ELIXR model. Using HOPPR's foundation model, they also showed the potential of automated reporting of not only the current CXR but also making comparisons with the prior exam in the style of the reporting radiologist. Other companies like Aidoc, Lunit Cancer Screening, and GLEAMER have shared their intent to leverage foundation models. In recent months, several papers, such as Med-Gemini, MedVersa, and Merlin, have reported the potential for automated CT reporting capabilities.
🚧 Taking automated reporting from research to actual clinical practice requires much more than just AI models. You also need to address issues related to regulatory, medicolegal, workflow, automation bias, ethics, clinical expertise required, etc. At the same time, the increasing workload, worsening global shortage of radiologists, and persistent staff burnout will motivate continued efforts in this area, both research and commercial.
🔗 Links to all the sources mentioned are in the first comment. #radiology #RadiologyAI #RadiologyReporting
