Curated benchmark dataset for ultrasound based breast lesion analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate breast lesion detection and classification are crucial for effective breast cancer diagnosis.
• Existing datasets may lack comprehensive annotations or patient-level data for robust AI model training.
• Breast ultrasound is a widely used imaging modality for breast lesion assessment.

Purpose of the Study:

• To introduce the Breast Ultrasound Scans (BrEaST) dataset, a novel resource for AI-driven breast cancer research.
• To provide a comprehensive dataset with detailed annotations for training and validating AI models for breast lesion detection, segmentation, and classification.
• To facilitate advancements in automated breast ultrasound analysis.

Main Methods:

• The BrEaST dataset comprises 256 breast ultrasound scans from 256 patients.
• Scans include images of benign lesions, malignant lesions, and normal breast tissue.
• Expert radiologists manually annotated images, providing BIRADS lexicon labels and histopathological classifications where available.

Main Results:

• The BrEaST dataset is the first of its kind to offer patient-level, image-level, and tumor-level labels.
• All cases are confirmed by follow-up care or core needle biopsy, ensuring data accuracy.
• The dataset contains freehand annotations for tumor identification and BIRADS feature labeling.

Conclusions:

• The BrEaST dataset provides a valuable, publicly accessible resource for AI research in breast ultrasound.
• This dataset will support the development of improved AI tools for breast cancer detection and characterization.
• Availability under CC-BY 4.0 license promotes collaborative research in medical imaging analysis.