Rectal Carcinoma Segmentation Based On Image Pre-processing and Improved U-Net with Residual Block

Abstract

Accurate staging of rectal carcinoma is crucial for surgery plan and the accuracy is mostly depending on the segmentation algorithm for CT images. There have been a lot of promising recent works utilizing deep network architectures, such as Full Convolutional Networks (FCNs) and If-Nets, for developing automated medical image segmentation, and there are few proposed for rectal carcinoma segmentation. A new effective image pre-processing method combined with CT window technique and image cropping is proposed in this paper and the proposed model established is an encoder-decoder convolutional network. The residual block instead of the convolutional block was used for U-Nets. Besides, the image augmentation technique was also utilized in this paper and the results demonstrated that the performance of segmentation can be greatly improved. The Dice Similarity Coefficient (DSC) reaches at 0.974 in validation set which exceeds most of the latest work on rectal carcinoma segmentation and it can be concluded that the proposed preprocessing technique on U-Nets with residual block can greatly improve the overall performance of rectal carcinoma segmentation.