A Competition, Benchmark, Code, and Data for Using Artificial Intelligence to Detect Lesions in Digital Breast...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate artificial intelligence (AI) for digital breast tomosynthesis (DBT) can enhance cancer detection and reduce global healthcare costs.
• Developing AI for DBT analysis requires accessible training data, clear benchmarks, and shared code.

Purpose of the Study:

• To create a benchmark for AI algorithms detecting lesions in DBT.
• To make training and evaluation data publicly available.
• To release code for existing AI methods in DBT analysis.

Main Methods:

• A multi-institutional international grand challenge was conducted.
• Research teams developed AI algorithms to detect lesions in DBT.
• A dataset of 22,032 DBT volumes was provided for algorithm development and testing across two phases.

Main Results:

• Eight teams participated, developing AI algorithms for DBT lesion detection.
• The top-performing team achieved a mean sensitivity of 0.957 for biopsied lesions.
• Aggregated mean sensitivity across all algorithms was 0.879, and 0.926 for Phase 2 participants.

Conclusions:

• An international competition yielded AI algorithms with high sensitivity for detecting DBT lesions.
• A standardized performance benchmark and publicly available data/code were released.
• These resources aim to accelerate future research in computer-assisted DBT diagnosis.