Multi-Modality Breast MRI Segmentation Using NNU-NET For Preoperative Planning Of Robotic Surgery Navigation

Abstract

Segmentation of the chest region and breast tissues is essential for surgery planning and navigation. This paper proposes the foundation for preoperative segmentation based on two cascaded architectures of deep neural networks (DNN) based on the state-of-the-art nnU-Net. Additionally, this study introduces a polyvinyl alcohol cryogel (PVA-C) breast phantom based on the segmentation of the DNN automated approach, enabling the experiments of navigation systems for robotic breast surgery. Multi-modality breast MRI datasets of T2W and STIR images were acquired from 10 patients. Segmentation evaluation utilized the Dice Similarity Coefficient (DSC), segmentation accuracy, sensitivity, and specificity. First, a single class labeling was used to segment the breast region. Then it was employed as an input for three-class labeling to segment fat, fibroglandular (FGT) tissues, and tumorous lesions. The first architecture has a 0.95 DSC, while the second has a 0.95, 0.83, and 0.41 for fat, FGT, and tumor classes, respectively.