Summary
Highlights
The Canon Medical Academy provides training and educational resources for medical professionals, ensuring they are equipped with the latest knowledge and skills. This webinar, hosted by Canon Medical, focuses on optimizing radiologist and AI collaboration in practice, featuring a diverse panel of experts.
The webinar introduces three leading experts in AI and radiology: Professor Kenji Suzuki, Professor Prrenav Rajpar, and Professor Curtis Langlots. The discussion sets the context by highlighting the rapid growth of FDA-approved AI devices, with radiology being the most AI-invested medical specialty. AI's role has expanded beyond diagnostic aid to non-interpretive tasks, and there's a shift towards multimodal AI and foundation models.
Despite advancements, concerns remain about AI acting alone due to flawed metrics, hallucinations, brittle performance, and a lack of deep clinical reasoning. The importance of humans in the loop, from design to post-deployment monitoring, is emphasized, likening it to the need for human supervision in self-checkout systems for age-restricted purchases.
A key discussion point is why clinicians hesitate to trust AI, despite its increasing use and FDA approvals. Dr. Langlots explains that AI systems need to be 'worthy of trust' by explaining their conclusions, providing confidence levels, undergoing clinical trials, and offering transparency in their training data. Professor Rajpar notes the disparity in clinical trials between radiology and other specialties, highlighting a study where AI improved some radiologists' performance while decreasing others', averaging to zero overall improvement. Professor Suzuki emphasizes the need for 'explainable AI' (XAI) that can provide reasoned bases for individual AI decisions in natural language.
The panel explores optimal AI-human collaboration models. Dr. Langlots suggests research into how AI and humans can best work together, mentioning a study where a chatbot comparing human and AI diagnoses significantly improved human performance. He also discusses the varying need for clinical trials based on the risk level of the AI application, from low-risk measurements to high-risk screening solutions.
The conversation shifts to barriers in generalizing AI-assisted radiology. Dr. Langlots identifies two primary modes: building algorithms for specific practices that don't require generalizability, and foundation models trained on massive, multi-institutional datasets for many different downstream tasks. Professor Suzuki discusses the potential of 'small data AI' for rare diseases, suggesting an optimal solution might be a combination of foundation models for generalists and small data AI for specialists. Professor Rajpar highlights the vast number of findings in common exams like abdomen pelvis CTs, where a foundation model approach is crucial for managing such complexity and could lead to significant productivity and accuracy gains, as seen in a recent draft report generation tool.
The panelists discuss how to make AI useful without adding to radiologists' burden. Dr. Langlots points out that current AI algorithms, often focused on triage and detection, can sometimes create extra work due to false positives. He emphasizes the excitement around report drafting tools, which can significantly improve efficiency and reduce burnout, drawing parallels to the impact of PACS. Professor Suzuki proposes a 'silent safety net AI' that only alerts for difficult cases, minimizing distraction and requiring extremely low false-positive rates. Professor Rajpar notes that the lack of comprehensive AI tools for common exams still limits adoption, but a shift in mindset among radiologists is occurring, driven by the desire for AI to handle routine tasks.
The discussion delves into how AI will change the radiologist's role. Dr. Langlots reflects on how PACS revolutionized radiology in unforeseen ways, suggesting AI will not displace radiologists but help manage increasing workloads, leading to more efficiency or improved work-life balance. Professor Rajpar contends that AI will eventually surpass human diagnostic image interpretation, leading to a redefinition of human roles to focus on aspects where human intelligence is uniquely valuable, such as patient interaction and communication with referring physicians. Dr. Langlots reinforces that human and AI intelligence are different, with humans excelling in context and relationships, and emphasizes that in high-risk situations like healthcare, human oversight remains crucial. He highlights AI's potential in prediction by analyzing rich, multimodal datasets (genomics, EHR, imaging).
The panelists project into the future, sharing what excites them most. Professor Rajpar uses the analogy of self-driving cars (Whimo) to illustrate how initial hesitation gives way to transformative experience once AI is trusted. He envisions an AI-powered reading room where AI determines which cases a radiologist should read (for intellectual stimulation or workload balance), generates reports, and handles measurements, making work more enjoyable. Professor Suzuki foresees an AI that acts as a 'colleague,' only alerting radiologists when necessary for complex cases. Dr. Langlots cautions about understanding AI's limitations, drawing a parallel to airplane autopilots—humans are needed for unexpected situations. He also considers areas where AI can 'run free' with guardrails, such as report drafting, where human supervision ensures accuracy.
The Q&A segment begins with a question about merging voice AI and large vision models for radiology. Dr. Langlots notes existing products that create structured reports from dictation and predicts a future with integrated cloud-based diagnostic workstations that combine image display, worklist, AI, and voice. Professor Suzuki warns about AI 'hallucinations' in report drafting, emphasizing careful review. Another question addresses the 'fate of radiologists,' with Dr. Langlots sharing research predicting a one-third reduction in hours needed for radiologists due to AI in five years, but this will likely be absorbed by increasing volumes rather than job displacement. The panel then tackles concerns about 'cognitive blunting' due to automation. Professor Rajpar acknowledges this trend, referencing how summaries and grammar tools reduce the need for deep engagement, and suggests accepting areas where skills can be offloaded to AI. Dr. Langlots stresses the radiologist's responsibility for signed reports and the need for education on AI's workings and potential failings to prevent complacency.
The final question explores patient involvement in AI decision models and human confidence when stakes are high. Dr. Langlots explains that while patients haven't historically needed notification about AI use in medicine, they want healthcare organizations to make good, transparent decisions about AI implementation, especially concerning safety and data privacy. Professor Rajpar notes a growing patient desire for direct access to and choice over AI technology, citing examples like preventative health scans. The webinar concludes with the hosts thanking the panelists and audience, reiterating that while AI won't replace doctors, doctors who don't embrace AI may be replaced by those who do.