Tumor segmentation of breast tomosynthesis in breast-conserving surgery with deep learning

Breast cancer stands as one of the most prevalent cancers affecting women globally. Early-stage breast cancer often undergoes breast-conserving therapy (BCT) coupled with irradiation, the primary treatment choice. This approach, retaining the breast through surgery, not only preserves its appearance but also minimizes the risk of recurrence. However, a positive margin following BCT can elevate the risk of local recurrences in any malignant tumor. Hence, reducing positive margins could provide surgeons with real-time intra-operative insights into resection margins. This study endeavors to formulate an intra-operative tumor segmentation utilizing breast tomosynthesis during breast-conserving surgery. Leveraging a practical deep learning model, i.e. U-net, to delineate the tumor margin. This work equips surgeons with enhanced information for achieving clean margins during breast-conserving surgeries. Evaluating 35 cases, this study compared its findings with manually determined contours and pathology reports. The experimental outcomes exhibited promising results, showcasing Intersection over Union (IOU) and Dice coefficients at 0.874 and 0.931, respectively. Through the integration of deep learning techniques, this proposed scheme could potentially revolutionize the intra-operative measurement system.