From Localization to Resection: Development and Validation of a Real-Time AI Navigation System for Vacuum-Assisted Breast Biopsy

Xinran Shao¹Yunzhi Shen²Pingdong Sun³Yihan Sun⁴Xingai Ju⁵Hongjie Zhu⁶Hong Li⁶Qiushi Li³Ruan Ting⁷Jinrui Liu³Yuqing Wang³Qikun Guo⁸Yuxin Ma⁹Xiang Fei¹Hang Sun¹⁰Cui Jianchun^3*

• ¹Department of Thyroid and Breast Surgery , People's Hospital of China Medical University, Shenyang, China
• ²the graduate school of dalian medical university, Shenyang, China
• ³Department of Thyroid and Breast Surgery, People's Hospital of China Medical University, Shenyang, China
• ⁴Department of Cardiology , People's Hospital of China Medical University, Shenyang, China
• ⁵Department of General Medicine, People's Hospital of China Medical University, Shenyang, China
• ⁶Northeastern University College of Medicine and Biological Information Engineering, Shenyang, China
• ⁷Liaoning University of Traditional Chinese Medicine, Shenyang, China
• ⁸China Medical University School of Artificial Intelligence, Shenyang, China
• ⁹China Medical University School of Artificial Intelligence, Shenyang, China
• ¹⁰School of Information Science and Engineering, Shenyang Ligong University, Shenyang, China

Vacuum-assisted breast biopsy (VABB) is a widely used, minimally invasive technique for diagnosing and treating breast tumors. However, the procedure requires precise real-time localization of both the tumor and knife under ultrasound guidance, which can be challenging for junior surgeons with limited radiological experience. To address this issue, we developed a novel two-stage real-time artificial intelligence (AI) navigation system based on the YOLOv11 deep learning architecture. Model 1 identifies the tumor and cutter slot, while Model 2 tracks the needle and tumor during the resection process. The system was trained and validated using 22,278 annotated ultrasound images collected from 167 clinical VABB procedures. Model performance was assessed using three-fold cross-validation and compared with the performance of junior surgeons. The AI system achieved significantly higher localization accuracy across all metrics. For Model 1, the mAP50 for tumor and cutter slot localization was 0.907 and 0.671, respectively. For Model 2, the mAP50 for tumor and knife tracking was 0.829 and 0.765. Inference speed was 1.2 ms per frame on GPU and 32.6 ms on CPU. These results demonstrate that the proposed AI system offers reliable, real-time guidance and holds strong potential to support junior surgeons and improve outcomes in clinical VABB procedures.