RadGame: AI-Powered Radiology Training

• RadGame is an AI-powered platform that gamifies radiology training by using large annotated datasets to deliver structured feedback on imaging localization and report writing.
• It employs state-of-the-art vision-language models and automated metrics (IoU and CRIMSON) to quantitatively assess radiologic skills, leading to significant learning gains.
• The framework scales personalized education through interactive modules that simulate real clinical workflows and are validated by prospective user studies.

RadGame is an AI-powered, gamified educational platform for radiology, designed to teach and assess two core radiologic competencies: localization of abnormalities on imaging and structured radiology report generation. The platform repurposes large, annotated public datasets of chest X-rays and leverages automated, model-driven feedback to provide immediate, personalized, and structured guidance at scale. RadGame utilizes state-of-the-art vision-LLMs and report evaluation metrics to deliver quantifiable improvements in diagnostic accuracy and reporting skill when compared to conventional passive learning methods (Baharoon et al., 16 Sep 2025).

1. System Architecture and Workflow

RadGame organizes the educational process into two interactive modules: RadGame Localize and RadGame Report.

• RadGame Localize presents users with chest X-rays and prompts them to identify abnormalities either by drawing bounding boxes (“Draw Findings”) or by selecting items from a findings checklist (“Select Findings”). These responses are quantitatively assessed against reference annotations provided by expert radiologists from public datasets.
• RadGame Report tasks users with composing narrative radiology reports, given a chest X-ray plus accompanying meta-information (patient age, clinical indication). The written reports are evaluated for correctness, coverage, and style via automated system feedback.

Gamification is central to the RadGame paradigm: both modules incorporate scoring, timed challenges, and progression-based case selection to drive engagement and accelerate learning. Gamified features convert the passive exposure typical of classic radiology education into an active, feedback-rich environment that closely simulates the interpretive tasks required in real clinical workflows.

2. Core Competencies and User Tasks

RadGame is engineered to target two principal skill domains:

• Localization Accuracy: Users must pinpoint the precise anatomical location of radiographic findings. This is operationalized by bounding box annotation with subsequent comparison to gold-standard radiologist-drawn boxes from open datasets. Statistical accuracy is measured using the intersection-over-union (IoU) metric, with a match threshold of 0.25 as established via expert consensus.
• Report Generation Skills: Users craft free-text findings (and, when prompted, impressions or summary sections) using their interpretation of the presented case. Reports are evaluated for factual accuracy (presence/omission of findings, correct location/severity, false positive/negative statements) and stylistic features (organization, completeness, use of clinical language).

Both modules foster transition from observational “passive review” to active interpretation and synthesis, aligning with the practical requirements of radiology training.

3. Automated Feedback and Evaluation Metrics

Immediate, structured AI-driven feedback is a cornerstone of RadGame.

• Localization Feedback: User-submitted annotations are evaluated using IoU computed against reference annotations. For each missed or mis-localized abnormality (IoU < 0.25), the platform produces a concise visual explanation using the MedGemma 4B vision-LLM. The explanation identifies the specific missed finding and offers a two-sentence guide to its salient visual features. This mechanism guides learners toward improved spatial pattern recognition.
• Reporting Feedback: The system employs a custom metric, CRIMSON, which extends the previously published GREEN metric. CRIMSON focuses exclusively on abnormal findings (deliberately ignoring correctly reported normal findings to avoid inflating scores) and incorporates patient age and indication to weight errors for clinical significance. Report assessment categorizes discrepancies into four error types: (a) false positives (claimed but absent findings), (b) missed findings, (c) mislocalization, and (d) misclassification of severity. The principal CRIMSON score is computed as:

$$\text{Score} = \frac{\text{Number of matched findings}}{\text{Number of matched findings} + \Sigma\text{(errors from categories a–d)}}$$

Automated report evaluation is performed by GPT-o3, which compares the user’s report to expert-generated references and generates a breakdown of errors, alongside a “Style Score” for linguistic and syntactic features such as logical order, section completeness (lungs, heart, mediastinum, bones), and sentence structure.

4. Quantitative Evaluation and Efficacy

A multi-institutional, prospective user paper measured the efficacy of RadGame relative to traditional passive-learning modalities:

• Localization Module: Users in the gamified cohort achieved a 68% improvement in post-test localization accuracy versus only 17% improvement with traditional methods. Additionally, time-per-case systematically decreased, indicating increasing efficiency with skill acquisition.
• Report-Writing Module: The gamified cohort showed a 31% pre-post improvement in CRIMSON scores, compared to a 4% gain in the passive group. Statistical significance was clearest for localization, though trends also favored the interactive approach for report-writing.

These findings demonstrate that RadGame’s AI-feedback-based gamification substantially magnifies training gains in both core skill domains within a fixed number of cases, compared to observation-only regimens.

5. Technical Implementation Details

RadGame utilizes several technical innovations:

• Image Annotation Scoring: Detection accuracy is established via IoU, with a 0.25 success threshold. Each annotation is linked to a specific finding label, mirroring real-world clinical annotation standards.
• Report Evaluation Pipeline: The CRIMSON metric is computed by parsing structured findings from generated and reference reports, weighted by clinical context. Assessment is automated by a LLM (GPT-o3), which constructs a tabulated error breakdown and generates prose feedback according to the four error types.
• Explanation Generation: MedGemma 4B, as a foundation vision-LLM, produces concise, context-specific explanations for missed findings. This provides immediate and actionable remediation.
• Style Assessment: Reports are scored for stylistic elements, including proper use of sections, sentence completeness, and adherence to clinical language conventions.

This pipeline permits fully-automated, scalable, and context-aware feedback to a diverse and large cohort of trainees.

6. Implications and Future Directions

RadGame exemplifies the application of AI and gamification to mediate scalable, interactive, feedback-intensive education in radiology. Key implications include:

• Enhanced and Quantifiable Learning: The platform achieves statistically robust improvements in learner competence, both in localization and reporting, with efficiency gains noted in reduced time per case.
• Scalability and Personalization: Automated feedback on public datasets obviates the need for direct faculty supervision and customizes remediation to each learner’s performance.
• Metric and Model Refinement via Human-in-the-Loop: The iterative comparison of user and model outputs (as facilitated by metrics such as CRIMSON) enables identification and remediation of both human and AI errors, serving as a de facto evaluation platform for radiology AI models.
• Extensibility: RadGame’s modular framework suggests extension to new imaging modalities (e.g., CT) and incorporation of dialogue-based (conversational) feedback, allowing for further alignment with evolving educational paradigms.

This suggests that RadGame’s paradigm—AI-driven gamified education with structured feedback—can be adapted to other domains within medicine and beyond, wherever both interpretative accuracy and structured communication are core competencies.

In summary, RadGame represents a comprehensive, data-driven approach to radiology education, integrating advanced AI models with interactive learning mechanisms to deliver immediate, tailored, and context-sensitive feedback (Baharoon et al., 16 Sep 2025). Its demonstrated improvements in fundamental skills highlight the transformative potential of AI in scaling and personalizing medical education.