Segmentation of Venous Vessel in MRI using Transferred Convolutional Neural Network

Abstract

Accurate segmentation of pancreatic venous vessels (PVV) is of great significance for clinical pancreatic cancer radiotherapy. Magnetic resonance imaging (MRI) can provide qualitative assessment of PVV through visualization. Therefore, there is a strong need for an automated segmentation method that can be directly applied to the segmentation of PVV in MR images. In this paper, we develop a deep learning based method with distance and contour regularized level set evolution (DCRLSE) model refinement for automatic detection and segmentation of PVV. The proposed method consists of two main steps: (1) venous vessel localization and segmentation using a dual convolutional neural network (DualCNN) (2) refinement of the initial segmentation with DLRSE model. As PBV is very challenging to segment due to the high inter-patient anatomical variability in both shape and size, the learned weights from other cases are utilized and finetuned to be applied on new cases to obtain consistent and accurate segmentation results. The proposed method was evaluated on 40 MRI scans from a clinical thoracic-abdomen dynamic MRI dataset. Dice similarity coefficient (DSC), sensitivity, specificity, and modified Hausdorff distance (MHD) were calculated to evaluate the segmentation performance of the proposed method. An average sensitivity: 87.60%, specificity: 99.60%, DSC: 86.3%, and MHD: 0.0991mm were obtained. Comparison with other state-of-the-art segmentation methods demonstrate that the proposed method can provide more accurate and consistent segmentation results of PVV in MR images.