Tencent's AI Breast Cancer Screening System: Technical Architecture and Implementation

This article presents Tencent's AI breast cancer screening system (Tencent Miying AI Breast System) developed for early breast cancer detection. The system addresses the growing demand for breast cancer screening in China, where the five-year survival rate (83%) lags behind the United States (89%) primarily due to a shortage of experienced radiologists.

The system covers multiple diagnostic modalities: mammography (钼靶), pathology, MRI, and ultrasound, with mammography being the most mature implementation deployed in over 30 tertiary hospitals. The mammography system achieves three main functions: lesion localization, malignancy probability assessment, and automated report generation.

The technical framework consists of three layers: preprocessing for mammography image normalization and device adaptation, AI learning models as the core module, and dynamic updates based on physician feedback. The TMuNet model employs four key innovations:

1. Four-image input (MLO-CC views for both breasts) enabling cross-comparison to solve "same shadow different disease" problems;

2. Multi-scale network design preserving lesion size information critical for malignancy determination;

3. Progressive network construction mimicking human learning - solving simpler problems first before tackling complex diagnoses;

4. Self-paced training starting with easier samples and gradually introducing harder cases.

Performance metrics show 92% sensitivity at 0.2 false positive rate for mass detection, and 87% sensitivity with 96% specificity for benign-malignant classification. For pathology, the mitosis detection system achieves F1 score of 0.82, surpassing the academic competition winner's 0.73. The system also developed a semi-automatic annotation tool for MRI lesion segmentation, significantly reducing physician annotation time from 30-60 minutes per image.

The ultimate goal is to integrate these multi-modal data sources to provide comprehensive diagnostic assistance for patients and physicians.