Research on the Whole-Process Intelligent Diagnosis and Treatment of Digital Breast Tomosynthesis Based on Deep Learning

Research on the Whole-Process Intelligent Diagnosis and Treatment of Digital Breast Tomosynthesis Based on Deep Learning: Multicenter Retrospective and Prospective Validation

This study aims to construct a multi-task deep learning model system to mine deep features in DBT images, so as to achieve accurate detection of breast lesions, differential diagnosis of benign and malignant (especially for the challenging BI-RADS 4A category), prediction of molecular subtypes, and evaluation of neoadjuvant chemotherapy (NAC) efficacy, providing an imaging basis for precision medicine.

Inclusion Criteria

1. Female patients aged ≥ 18 years.
2. Complete bilateral digital breast tomosynthesis (DBT) images available, including craniocaudal (CC) and mediolateral oblique (MLO) views.
3. Confirmed pathological diagnosis (core needle biopsy or surgical resection) serving as the reference standard; or benign lesions with stable findings on follow-up for more than 2 years.
4. (For the efficacy prediction subgroup) Patients who received complete neoadjuvant therapy and had postoperative pathological results.

2. Exclusion Criteria

1. Poor image quality with severe artifacts that precluded reliable analysis.
2. History of previous breast surgery or radiotherapy (except for the recurrence risk subgroup).
3. Incomplete clinical or pathological data.