Automated Vessel Segmentation in Pancreas 4D-MRI using a Novel Transferred Convolutional Neural Network

Abstract

Pancreatic cancer has dismal prognosis with five-year survival of ∼5%. Although surgery is considered the most effective treatment, only 10-20% of the patients present with resectable disease. Stereotactic body radiation therapy (SBRT) with simultaneous integrated boost (SIB) to tumor sub-volumes abutting major vessels has the potential to sterilize the tumor sub-volumes and down-stage patients for subsequent successful resection. However, abdominal tumor and organs are mobile. SBRT-SIB to small tumor sub-volumes is greatly challenged by ventilation induced tumor and organ motion, making it difficult to define treatment margins. 4D imaging has been used to assess tumor motion in radiation therapy. In a 4D image set, 5 to 10 3D volumetric images are usually generated, making manual segmentation of involved vessels time-consuming and not practical in real clinical setting. The purpose of this study is to apply a novel deep-learning method, two-scale convolutional neural network (CNN) followed by distance and contour regularized level set evolution (DCRLSE), on pancreas 4D-MR data sets to automatically delineate vessels limiting surgery. The 4D-MR images of ten patients were used. Each patient has 10 3D respiration-resolved image sets. For each 3D image set, there are approximately 40 2D axial slices that include vessels limiting surgery. For each patient’s 4D-MRI data-set, 300 image frames were analyzed. The proposed transferred CNN routine is composed of two tasks. In the source task, a relatively large amount of data is used to train the paired CNN. In the target task, a relatively small amount of data is used to fine-tune the network. In the test phase, Paired CNN is used to locate and roughly segment the vessel. Small holes in the segmentation are filled using morphological operation. The DCRLSE model is then used to refine the contours. Transferred learning is used to improve the performance by exploiting the similarity among different patients. After the binary mask is trained on one patient, the network weights are transferred to initialize the CNN training of the second patient’s network. The weights are fine-tuned on selected breathing phases to find the minimal number of images needed for the new segmentation task. A widely used segmentation method, level-set, was used for comparison. Manual segmentation was used as ground truth. The segmentation results were evaluated visually, as well as quantitatively using four metrics: sensitivity, specificity, dice similarity coefficient (DSC) and modified Hausdorff distance (HD). CNN-DCRLSE method visually showed more consistent segmentation results comparing to level-set method. It was able to achieve average sensitivity of 0.876±0.070 (mean±σ), specificity of 0.999±0.001, DSC of 0.863±0.081, and HD (mm) of 0.099±0.059. The CNN-DCRLSE method was successfully implemented on 4D-MR images to automatically segment major blood vessels that limit surgery for pancreatic cancer patients.