ISCAP Proceedings: Abstract Presentation
Lessons Learned from Designing an AI-Supported Radiology Education Platform
Jennifer Schulte
Dakota State University
Abstract
Demand for self-directed radiology education platforms is growing, but design guidance for integrating AI components into these tools remains thin. ASPIRE (AI-Supported Platform for Innovative Radiology Education) is a design science research artifact: a Streamlit application combining spaced repetition, visual localization overlays, the CheXlocalize dataset, and LLM-driven tutoring via Mistral-7B-Instruct-v0.3. Three computational studies conducted without human participants evaluated the platform's visual feedback and tutoring components. Construction of the artifact yielded three lessons.
First, model explainability is not a pedagogical explanation. Across 1,710 image-finding pairs, Grad-CAM saliency maps achieved a mean IoU of 0.102 and a hit rate of 0.187 against expert reference segmentations, far below what a learner-facing localization cue requires. Second, the simpler annotation won; across 607 paired comparisons, expert bounding boxes reached a mean IoU of 0.529 and outperformed Grad-CAM masks on all ten findings, indicating that curated human annotation, rather than derived model attention, should anchor visual feedback. Third, across 160 cases and four prompting strategies, role-conditioned and direct prompts scored highest on clinical accuracy and pedagogical appropriateness, while Socratic prompting, the strategy most aligned with established tutoring practice, underperformed. This suggests that a 7B-parameter model's instruction-following capacity, not pedagogical theory, constrains prompt design at this scale.