Uncertainty-Aware Breast Ultrasound Explanation Cards: A Visual Communication Framework for Image-Based AI Diagnostic Support Using BreastMNIST_224

AI-assisted diagnostic interfaces should communicate more than a class label. They need to show the predicted risk, the uncertainty around that risk, the visual evidence that influenced the model, the limits of the evidence, and the appropriate next action. This paper presents an uncertainty-aware explanation-card framework for breast ultrasound decision-support screens. The empirical study was conducted on BreastMNIST_224, the 224 x 224 MedMNIST+ breast ultrasound benchmark with official train, validation, and test splits of 546, 78, and 156 images. The positive class was defined as malignant. Five image classifiers were trained on downsampled image grids, and the selected card model was a Platt-probability RBF SVM. On the official test split, the selected model achieved AUROC = 0.867 and AUPRC = 0.728. A validation-selected operating threshold of 0.254 gave accuracy = 0.769, sensitivity = 0.833, specificity = 0.746, Brier score = 0.125, and ECE = 0.068. The explanation card pairs malignant-risk probability with risk tier, uncertainty band, occlusion-sensitivity heatmap evidence, a limitation statement, and a review cue. In the held-out test set, the conservative Low-risk tier contained six cases and no malignant cases; all seven false negatives occurred in the Review tier rather than in Low risk. These findings support a prototype-level visual communication framework in which image evidence is shown together with uncertainty and safeguards, while diagnostic authority remains with the clinician.