Learning Breast Tissue Prone-To-Supine Displacement for Surgical Planning with Convolutional Neural Networks

Breast-conserving surgery is the preferred treatment for non-palpable early-stage breast tumors. MRI images in the prone position are commonly used for accurate diagnosis of lesions. However, as the procedure is performed in the supine position, significant breast deformation occurs. Therefore, preoperative localization of the lesions is necessary to ensure their effective removal during surgery. In this study, we propose a novel deep learning based approach to predict the large deformation of the preoperative volume of the breast given the intraoperative surface. We use a fully connected network trained in a supervised manner using synthetic generated ground-truth displacement vector fields. We evaluated the method in 10 real clinical cases with an average tumor localization error of 15.46 ± 3.96 mm and an average tumor-skin projection distance of 13.38 ± 4.61 mm.