Multiscale receptive field based on residual network for pancreas segmentation in CT images

Medical image segmentation has made great achievements. Yet pancreas is a challenging abdominal organ to segment due to the high inter-patient anatomical variability in both shape and volume metrics. The UNet often suffers from pancreas over-segmentation, under-segmentation and shape inconsistency between the predicted result and ground truth. We consider the UNet can not extract more deepen features and rich semantic information which can not distinguish the regions between pancreas and background. From this point, we proposed three cross-domain information fusion strategies to solve above three problems. The first strategy named skip network can efficiently restrain the over-segmentation through cross-domain connection. The second strategy named residual network mainly seeks to solve the under- and over- segmentation problem by cross-domain connecting on a small scale. The third multiscale cross-domain information fusion strategy named multiscale residual network added multiscale convolution operation on second strategy which can learn more accurate pancreas shape and restrain over- and under- segmentation. We performed experiments on a dataset of 82 abdominal contrast-enhanced three dimension computed tomography (3D CT) scans from the National Institutes of Health Clinical Center using 4-fold cross-validation. We report 87.57 +/- 3.26 % of the mean Dice score, which outperforms the state-of-the-art method, producing 7.87 % improvement from the predicted result of original UNet. Our method is not only superior to the other established methods in terms of accuracy and robustness but can also effectively restrain pancreas over-segmentation, under-segmentation and shape inconsistency between the predicted result and ground truth. Our strategies prone to apply to clinical medicine. (C) 2019 Elsevier Ltd. All rights reserved.