Data Sheet 1_Optimizing breast cancer ultrasound diagnosis: a comparative study of AI model performance and image resolution.docx

ObjectivesTo determine the optimal combination of artificial intelligence (AI) models and ultrasound (US) image resolutions for breast cancer diagnosis and evaluate whether this combination surpasses the diagnostic accuracy of senior radiologists.
Materials and methodsWe systematically compared lightweight (MobileNet, Xception) and dense neural networks (ResNet50, DenseNet121) using three image resolutions (224 × 224, 320 × 320, 448 × 448 pixels). A retrospective cohort of 4,998 patients was divided into training/validation (8:2 ratio, n = 3,578) and independent testing sets (n = 1,410). Diagnostic performance was assessed via AUC, sensitivity, specificity, and analysis speed, with direct comparisons against senior radiologists.
ResultsMobileNet with 224 × 224 input achieved the highest AUC (0.924, 95% CI: 0.910–0.938) and accuracy (87.3%) outperforming senior US (AUC: 0.820, accuracy: 79.1%) and mammography doctors (AUC: 0.819, accuracy: 83.6%) (p < 0.05). After excluding BI-RADS 4c and 5 nodules, the diagnostic efficacy of MobileNet_224 is better than that of senior doctors (p < 0.05), can reduce 60.1% false positives of US, and 46.6% of mammography. MobileNet_224 and MobileNet_320 had the fastest analysis speed.
ConclusionMobileNet_224 represents a novel, efficient AI framework for breast cancer diagnosis demonstrating superior accuracy and speed compared to both complex AI models and experienced clinicians. This work highlights the critical role of optimizing model architecture and resolution to enhance diagnostic workflows and reduce unnecessary biopsies.