乳房超音波画像データセットのためのコンピュータ支援診断

The Computer-Aided Diagnosis for Breast Ultrasound Imaging (CADBUSI) dataset comprises 79,281 breast ultrasound exams from 60,688 patients collected between 2002 and 2025 at Mayo Clinic Health System, specifically curated to advance machine learning applications in breast cancer diagnosis. This comprehensive collection includes 756,315 breast ultrasound images and 136,197 videos with BI-RADS® assessments and pathology-verified diagnoses, providing ground truth labels for 79,281 unique breasts (68,645 benign, 10,636 malignant). Each breast is classified based on the presence or absence of malignancy rather than individual lesion characterization, making this dataset particularly suited for multiple instance learning (MIL) approaches. Our rigorous processing pipeline standardizes the dataset while preserving clinical relevance through custom Faster R-CNN-based text extraction, automated detection of diagnostic regions, Noise2Noise inpainting for measurement caliper removal, and HIPAA-compliant anonymization. By addressing key challenges in ultrasound image standardization and linking radiological findings with pathological outcomes, this dataset enables the development of computer-aided diagnostic tools with the potential to improve breast cancer detection accuracy, reduce unnecessary biopsies, and enhance clinical decision-making. All preprocessing and dataset creation code is available at https://github.com/Poofy1/CADBUSI-Database . The dataset is not publicly available.