Ultrasound Video-Based Deep Learning Model for Predicting Axillary Lymph Node Status and Nodal Burden in Breast Cancer.

RATIONALE AND OBJECTIVES: Accurate preoperative assessment of axillary lymph node (ALN) status and nodal burden is crucial for individualized management of patients with breast cancer. This study aimed to develop and validate a two-stage deep learning (DL) framework that leverages preoperative breast ultrasound videos to predict ALN status and nodal burden. MATERIALS AND METHODS: In this multicenter retrospective study, 864 patients with pathologically confirmed breast cancer (July 2019-December 2024) were analyzed and divided into a training set (n=495), an internal test set (n=213), and two external test sets (n=120 and 36). A two-stage framework, based on a Temporal Shift Module (TSM) video model, was proposed to first predict ALN status (negative vs positive) and subsequently classify ALN-positive patients via nodal burden (1-2 vs ≥3 nodes). Model performance was evaluated using AUC, sensitivity, and specificity, along with subgroup analyses and comparisons with other DL and clinical models. RESULTS: For ALN status prediction, the TSM-ResNet50 yielded AUCs of 0.851 (95% CI, 0.803-0.894), 0.886, and 0.772 across the internal and two external test sets. Performance was consistent across key subgroups, including tumors >2 cm (0.870) and BI-RADS 4 C lesions (>0.875). For nodal burden prediction, the TSM-ResNet18 achieved AUCs of 0.937, 0.797, and 0.667 for internal and two external test sets, respectively. CONCLUSION: A two-stage video-based DL model showed promising performance in predicting ALN status and moderate yet clinically meaningful performance in predicting nodal burden, supporting its potential value for preoperative axillary assessment and individualized management.