Uncertainty-Aware Medical Image Explanation Cards: LLM-Generated Visual Explanations for AI-Assisted Radiology Interfaces

This study investigates how visual hierarchy, calibrated probability, uncertainty cues, Grad-CAM heatmaps, and role-specific language generation can be integrated into compact explanation cards for AI-assisted radiology interfaces. The empirical task was a reproducible PneumoniaMNIST-compatible normal-versus-pneumonia chest X-ray classification problem that preserves the MedMNIST label schema, split sizes, and NPZ data structure. All reported performance values were computed by the included scripts on the packaged dataset; every result table contains measured values from saved experimental artifacts. Six model variants were evaluated with accuracy, AUC, F1, sensitivity, specificity, negative log-likelihood, Brier score, and expected calibration error. The selected Spatial-CNN with temperature scaling achieved AUC = 0.868, accuracy = 0.763, F1 = 0.778, specificity = 0.923, Brier score = 0.155, and ECE = 0.021 on the 624-image test split. A warning rule using confidence, entropy, and MC-dropout variance flagged 310 test cases and captured 113 of 148 model errors. Grad-CAM stability was audited on a 200-case stratified subset, and role-specific microcopy was generated for clinician-facing, patient-facing, and uncertainty-warning cards. Patient-facing text achieved a mean Flesch Reading Ease of 74.6 and FK grade of 5.4, while clinician text preserved concise technical language. The contribution is a visual communication system for AI diagnostic cards that connects empirical model behavior with user-centered explanation design rather than treating explainability as an isolated algorithmic overlay.