Semantic segmentation of pancreatic medical images by using convolutional neural network

Abstract

Pancreatic cancer is the most difficult-to-detect cancer with the highest fatality rate. Pancreas analysis through abdominal computed tomography (CT) is challenging because the pancreas possesses a complex background and blurred boundaries with other organs. Traditionally, the notation of pancreatic area requires manual semantic segmentation by professional radiologists, but this is time-consuming. The developments of computer-assisted diagnosis of deep convolutional neural networks on pancreas segmentation are successful but require heavy computational complexity. This research is dedicated to the semantic segmentation of pancreatic CT images using convolutional neural networks to achieve favorable performance on pancreas segmentation with low computational parameters. We propose MobileNet-U-Net (MBU-Net) by combining U-Net model with light-weight MobileNet-V2 network. The proposed MBU-Net is assessed on the NIH pancreas-CT dataset. The averages for dice coefficient, Jaccard similarity coefficient, AUC, precision, specificity, and recall are 82.87%, 70.97%, 90.54%, 89.29%, 99.95%, and 77.37%, respectively. Results have demonstrated that the proposed MBU-Net can effectively reduce the computational parameters and achieve comparable performance when compared to the state-of-the-art methods.