Jonathan Smith

Fellowship Bio

Jonathan is a cardiology doctor training in the Wessex deanery with a subspecialty interest in heart failure and advanced echocardiography. He is passionate about how AI can revolutionise cardiovascular care and healthcare efficiency to deliver better patient care.

Fellowship Project

Before treating cancer with radiotherapy, clinicians must contour the tumour and nearby organs to protect using a planning CT scan. Commercial AI models can draw the organ contours, but clinicians must still check and correct these. This project was based at University Hospitals Southampton, where a locally trained AI was developed to act as an automatic contour quality checker. The idea is to predict whether the commercial AI’s outlines are good enough to use or need substantial edits, so that flagging only those needing attention spares clinicians the time spent re-checking everything. My role was to design and run two retrospective trials. The local model draws its own contours, and by measuring how closely the local and commercial outlines overlap, I tested whether that agreement could predict how closely the commercial outlines matched the final clinical ones. The trials differed in the reference standard used. One relied on contours drawn by clinicians from scratch, the other on commercial AI contours that clinicians then adjusted. The first predicted contour adequacy moderately well, the second poorly. That contrast was informative, as it shows clinicians contour differently from an AI draft than from a blank slate. I therefore set up retraining the model using 200 cases where clinicians definitively edited from the AI outlines, so the model learns the real workflow. I am now testing whether the model’s internal confidence in labelling each image point as an organ can predict contour adequacy. Next steps involve evaluating whether retraining or this confidence-based approach improves prediction. If it does, the aim is to extend to other anatomical regions and build the clinical safety and regulatory case needed to integrate the tool into an automated workflow safely.

Fellowship Testimonial

Graduating from this fellowship, what stands out most is how much broader my thinking about AI in healthcare has become. Coming in as a clinician, I’ve come away able to critically evaluate an AI tool and its uses within the NHS. That means weighing not just its technical performance, but the safety of putting it into practice and the very real technical and human barriers to deploying it. Working on the contour quality-checking project made that concrete. I learned as much from where the model struggled as from where it worked and now have a much clearer sense of AI’s limitations and the ways it can go wrong if not handled carefully. The monthly training days were a genuine highlight with each pairing excellent speakers with interactive workshops. They built a rounded understanding of AI in healthcare across model architecture, development, bias, regulation, deployment and clinical safety officer training. I also thoroughly enjoyed the process itself. My supervisors and the fellowship faculty were supportive throughout, and the cohort of fellows were a varied, engaging group who were a pleasure to learn from and alongside. For my future career, the fellowship has been formative. I now want to bring these skills back to where my clinical passion lies, in heart failure and echocardiography. Cardiovascular medicine is rich with imaging, physiology, and data, and I’m excited by how much AI could offer, from earlier detection of failing hearts to sharper echo interpretation. This fellowship has given me the technical literacy alongside the safety and governance grounding to pursue that responsibly. I leave determined to help build a future where AI genuinely improves the lives of cardiovascular patients.