Why Most Radiology AI Fails: Workflow, Governance & Integration Problem with Tessa Cook

Advanced engineering and deep learning can create breathtaking diagnostic models, but how do we successfully integrate complex AI tools into a high-volume, multi-site hospital system without creating more chaos? In this episode of Frame by Frame: Rethink Imaging, host Chris St. John sits down with Dr. Tessa Cook, a cardiovascular radiologist, national leader in imaging informatics, and Commission Chair for Informatics for the American College of Radiology (ACR), to break down the reality of clinical technology adoption. Dr. Cook explains the profound structural difference between a model that performs well in a retrospective testing lab and one that actually adds value to a radiologist’s busy shift. She shares concrete case studies detailing how her team built "Arnie" (Automated Radiology Recommendation Tracking Engine) to bridge fragmented patient care, ultimately giving rise to Penn Medicine's enterprise-wide High-Risk Follow-Through program that caught 12 early-stage cancers in its initial phase. We also look at the industry-wide hurdles to scaling medical AI, analyzing Penn's comprehensive three-phase evaluation framework, the evolution of clinical governance, and why standard technical metrics fail when algorithms completely lack access to crucial patient clinical context. What You’ll Learn: • The Workflow Commandment: Why the absolute highest-performing AI algorithm is completely useless if it cannot seamlessly connect to a radiology practice's existing daily workflow. • The Three-Phase AI Evaluation: How Penn Medicine rigorously tests new models—moving from retrospective performance metrics to limited prospective user experience testing before making a purchase. • The Clinical Context Gap: Why pixel-based AI tools experience high discordance rates when forced to evaluate medical imaging in complete isolation from a patient’s full medical history and lab results. • Closing the Follow-Up Loop: How mining structured data elements allows automated tracking engines to nudge clinicians and prevent critical downstream diagnoses from falling through the cracks. • Reimaging ROI in Healthcare: Why the return on investment for clinical technology must look beyond simple dollars to measure cognitive burden reduction, clinician burnout, and direct lives saved. . Chapters: • [00:00] The Core Informatics Problem: Why workflow integration dictates the success or failure of even the best clinical AI. • [02:00] Leadership Lessons from the Kitchen: An unexpected look at how Gordon Ramsay’s high-pressure team dynamics mirror healthcare operations. • [05:43] The Engineer-Physician Trajectory: Combining computer science, bioengineering, and medicine to tackle complex workflow obstacles. • [07:55] Defining Practice Transformation: Shining a light on systemic healthcare problems to optimize patient and clinician experiences. • [08:47] Going Back to 2010: Developing the open-source Radiance dose tracking software during residency to combat radiation overexposure. • [13:30] The Flaws of Pennsylvania Act 112: Analyzing the real-world operational challenges of mandated patient test notifications. • [16:38] Fragmented Care & Information Exchange: Why a lack of unified data sharing between healthcare systems allows critical follow-ups to get missed. • [18:10] Engineering "Arnie": Building an automated recommendation tracking engine using discrete data elements long before the LLM boom. • [20:10] 12 Lives Saved: Measuring the profound clinical success of Penn Medicine's enterprise-wide High-Risk Follow-Through program. • [21:23] Redefining the ROI of Healthcare AI: Looking past financial metrics to evaluate safety, efficiency, and cognitive burden reduction. • [23:58] Penn's 7-Year AI Governance Evolution: Streamlining rigorous validation models and achieving the ACR's ArchAI center designation. • [26:51] A Three-Phase Evaluation Strategy: Breaking down retrospective verification, limited prospective testing, and user experience deployment. • [28:53] Discordance vs. Ground Truth: Navigating the nuances of imaging measurements and the vital role of patient outcomes. • [30:11] The Missing Data Points: Why pixel-only AI algorithms fail to interpret imaging with the nuance of a specialized radiologist. • [32:52] Training the Next Generation: Graduating over 100 alumni from Penn’s elite Imaging Informatics Fellowship. • [37:16] An Apologetic Shift Toward AI: Balancing modern generative AI tools with foundational imaging standards like DICOM and HL7. • [39:23] The Value of Society Governance: Key takeaways from serving as Chair of the Society for Imaging Informatics in Medicine (SIIM). Frame by Frame: Rethink Imaging Podcast is handcrafted by our friends over at: fame.so (https://www.fame.so/)