Semantic Segmentation of Pancreas in Computed Tomography Images using Convolutional Neural Network

Diagnosis of pancreatic cancer is challenging due to complex background and fade boundaries in Computed Tomography (CT) images. Radiologists segment the pancreatic area manually by using semantics, which takes more time. This research aims to achieve a desirable performance on pancreatic segmentation with less computational power and effort through semantic segmentation using convolutional neural network. The proposed DeepLab-ResNet-SENet (DRSE-Net) is evaluated on National institutes of Health (NIH) pancreas CT dataset combining DeepLabV3 with ResNet101 and Squeeze-Excitation Blocks. Accordingly, obtained average Dice-Score, Jaccard Index, Precision and Recall are 0.8890, 0.7996, 0.8871, 0.8926 respectively compared to state of art approaches by reducing the parameters of the proposed model.