An interpretable ultrasound-based deep learning system for early breast cancer in a Chinese population
An AI breast ultrasound system explains its reasoning and helps radiologists detect cancer 6% more accurately, addressing how doctors can trust and use artificial intelligence in diagnosis.
BrcaDetect integrates ultrasound image features, BI-RADS scores, and demographics to support early breast cancer diagnosis, achieving AUC 0.989 in training and improving radiologist accuracy by 6% in a reader study across 5 hospitals. The model provides Grad-CAM and Shapley value explanations, addressing interpretability barriers that limit clinical adoption of AI diagnostics.
What the study was
- Study design
- Retrospective multicenter development + external validation cohort + reader study
- Population
- Women undergoing breast ultrasound evaluation across 5 hospitals in China (training + internal validation: 2399 patients; external validation: 649 women)
- Sample size
- 3048
- Category
- Early Detection
- Maturity
- Exploratory
- Journal
- Insights Imaging
Why it surfaced
HIGH by EARLY_CANCER_DETECTION flag (score 7 + flag). Multicenter external validation plus reader study distinguishes this from single-centre AI studies; interpretability features (Grad-CAM + Shapley) are direct response to clinical adoption barriers. Retrospective design and China-only cohort limit immediate generalisability.
A plain-language summary of published research — not medical advice. Talk to a clinician about your own care.