Construction of a Benchmark for Breast Ultrasound AI Interpretation and Performance Evaluation of Multimodal AI Models

Background: Breast cancer is the most prevalent malignancy among women worldwide. Ultrasound is a first-line screening modality, particularly in Asian populations with dense breast tissue where mammographic sensitivity is limited. However, ultrasound interpretation is highly operator-dependent, with substantial inter-observer variability in BI-RADS classification, especially for category 4A-4B lesions. Multimodal large language models (MLLMs) have emerged as a promising tool for medical image analysis due to their zero-shot diagnostic capability, interpretable chain-of-thought reasoning, and structured report generation. Nevertheless, there is currently no standardized benchmark for evaluating AI performance in breast ultrasound interpretation. Study Design: Approximately 1,380 breast ultrasound images will be curated (1,200 evaluation set + 150 out-of-distribution safety test set + 30 prompt development set), encompassing three diagnostic categories: normal breast, benign lesions (BI-RADS 2-4B), and malignant lesions (BI-RADS 3-5). Two junior radiologists (\<5 years of experience) and two senior radiologists (\>15 years) will independently annotate images per ACR BI-RADS v2025 with arbitration by a fifth expert for discordant cases. Diagnostic difficulty will be stratified into three tiers using cross-architecture deep learning consensus: Tier 1 (straightforward, both models correct), Tier 2 (equivocal, one correct/one incorrect), and Tier 3 (difficult, both incorrect, with senior expert validation). MLLMs will be evaluated across multiple dimensions: classification accuracy, sensitivity, specificity, F1 score, AUC, Cohen's kappa agreement with expert consensus, expected calibration error (ECE), morphological feature description accuracy, and chain-of-thought reasoning quality. Safety Assessment: (1) Out-of-distribution rejection test using 150 non-diagnostic images (degraded images, non-breast ultrasound, other imaging modalities); (2) Temperature-stability pre-experiment across parameter settings; (3) Thinking-mode ablation comparing standard vs. chain-of-thought reasoning modes. All experiments use fixed model snapshots, system fingerprint monitoring, and complete logging for reproducibility.