RadGame Localize: AI Radiology Training

• RadGame Localize is an AI-powered, gamified training module that teaches accurate abnormality localization on chest X-rays using quantitative IoU evaluation.
• It integrates vision-language models to deliver immediate, dual-mode visual and textual feedback, enhancing trainee spatial reasoning and diagnostic precision.
• The system improves localization accuracy and speed, showing a 68% improvement over traditional methods, thus driving more effective and scalable radiology training.

RadGame Localize is an AI-powered, gamified training module within the RadGame platform designed to cultivate the core radiology skill of abnormality localization on chest X-rays. The system integrates large-scale public datasets, automated feedback, and vision-LLMs to produce an interactive, accurate, and scalable educational environment. By requiring active engagement—bounding box annotation—RadGame Localize directly addresses the limitations of passive radiology case review, embedding immediate feedback and performance analytics into the learning process.

1. System Structure and Learning Objectives

RadGame Localize tasks participants with the identification and spatial delimitation of radiologically significant findings on chest X-rays. Each image presents a mix of "Draw Findings"—focal abnormalities requiring precise bounding boxes—and "Select Findings"—diffuse or anatomically fixed abnormalities selectable via checklist. The core pedagogical objective is to move beyond recognition towards proficient spatial localization, which is essential for clinical diagnosis and reporting.

Upon case presentation, trainees annotate findings with bounding boxes. These annotations are automatically compared against radiologist-generated ground truth boxes sourced from the PadChest-GR public dataset. The system implements a quantitative Intersection-over-Union (IoU) criterion: an annotated finding is deemed correct if the IoU, defined as

$$\mathrm{IoU} = \frac{\mathrm{Area\;of\;Overlap}}{\mathrm{Area\;of\;Union}}$$

exceeds a threshold of 0.25. This metric enables consistent, scalable evaluation of localization precision, aligning the trainee’s spatial reasoning with expert standards.

2. Integration of Gamification and AI Feedback

RadGame Localize operationalizes a gamified experience, transforming the localization exercise into a competitive and iterative game. Participants receive immediate feedback after each case, facilitated by the integration of vision-LLMs, specifically MedGemma 4B. When a localization mistake occurs—either a missed or incorrectly drawn bounding box—the system generates a concise, two-sentence visual explanation describing the radiographic features, anatomical context, and critical cues identifying the abnormality.

Feedback delivery is both visual and textual: the ground-truth annotation overlays the original image, and the generated explanation details why the abnormality should have been recognized. This dual approach is extended to "Select Findings," where descriptive feedback elucidates the diagnostic radiographic features even when only checklist selection is required. Such structured, AI-powered feedback enables trainees to iteratively refine their visual search strategies and diagnostic reasoning.

3. Localization Algorithm and Evaluation Metrics

The annotation comparison is executed by an automated algorithm that matches user-drawn bounding boxes against the dataset’s expert annotations. For each case:

• The algorithm computes the IoU between the trainee box and each ground-truth box in the annotated image.
• If IoU > 0.25 for any match, the localization is coded as correct.
• Missed findings (no sufficient overlap) trigger the vision-language feedback cycle.

The system’s evaluation metrics focus on both localization accuracy and speed:

• Improvement in correct localization rate, assessed before and after training.
• Reduction in time required per case, indicating increases in efficiency and search strategy optimization.

A prospective evaluation demonstrated a 68% localization accuracy improvement for Gamified RadGame users compared to 17% for traditional passive case review (statistically significant, two-tailed Mann–Whitney U test, p < 0.05). Additionally, participants showed an average per-case time reduction of 25 seconds.

4. Educational Impact and Clinical Significance

RadGame Localize advances radiology training beyond conventional paradigms by offering scalable, individualized, and feedback-rich education. The fusion of AI-powered annotation comparison and vision-LLM explanations provides structured guidance that can be tailored to each trainee.

This approach promotes active skill acquisition, improved retention of imaging findings, and enhanced spatial reasoning. The immediate corrective feedback accelerates the learning curve, enabling users to assimilate expert-level visual patterns and avoid repeated errors. A plausible implication is that such platforms may reduce the time required for trainees to achieve clinical proficiency in abnormality localization, potentially standardizing skills across institutions.

5. Challenges, Limitations, and Ongoing Development

While early evaluations indicate substantial performance gains, several limitations are noted:

• Balancing case interactivity and training efficiency requires ongoing tuning.
• AI-generated explanations must maintain clinical accuracy and specificity; errors or non-specific feedback may mislead trainees regarding diagnostic features.
• The localization threshold (IoU > 0.25) is sensitive and may require calibration according to case mix, trainee expertise, or specific educational goals.

Continuous updating and validation of the underlying AI components, including the vision-LLMs and algorithmic thresholds, are necessary to maintain clinical relevance and educational rigor.

6. Context within AI-Driven Medical Education

RadGame Localize exemplifies a new direction in radiology education, where automated performance analytics and interactive, immediate feedback serve as core pedagogical tools. The system leverages public radiology datasets and advanced multimodal AI to combine objective annotation comparison with qualitative, teaching-oriented explanations.

This methodology—gamified, feedback-rich, and scalable—signals a paradigmatic shift away from passive learning and toward data-driven, adaptive medical education platforms. These platforms may inform the structure and evaluation of future radiology curricula and extend applicability to other domains requiring spatial reasoning and expert-level pattern recognition.