Uncertainty-Aware Lightweight Breast Ultrasound Modeling on BreastMNIST+: Parameter-Efficient Label-Prompt Alignment, Controlled-Shift Robustness, and Verifiable Evidence Cards

Jiaqi Sun (The University of Texas at Austin)
Tyler Henderson (University of Washington)
Kevin Zhao (University of California San Diego)
Ashley Cooper (Purdue University System)

Journal of Information Systems and Business Technology
Vol 2 No 4 (2026): Journal of Information Systems and Business Technology

On BreastMNIST+ 224, we tested visual and class-prompt models with calibration, conformal sets, routing, perturbations, and evidence cards. The 546/78/156 train/validation/test artifact contained one conflicting duplicate. The ensemble achieved AUROC/AUPRC 0.885/0.738, balanced accuracy 0.842, sensitivity 0.833, and specificity 0.851. Temperature scaling reduced Brier/ECE 0.120/0.093-to-0.116/0.079. Label-conditional 90% conformal prediction achieved overall/malignant coverage 0.897/0.952 and size 1.359. Rank-2 prompts reached AUROC 0.790; blending raised sensitivity to 0.905 but lowered specificity to 0.763. Rotation and contrast degraded most. 156 cards passed checks. Results support routing, not clinical validity.