Abstract: Colorectal cancer (CRC) is among the top three tumors in the world. At present, screening CRC in CT images still needs to be done manually by radiologists, which is a time-consuming and laborious repetitive work. This paper investigated a method which extracted 3D features on abdominal CT images based on 3D convolutional network, then estimated whether there were CRC lesions through CT images. This paper collected a colorectal cancer CT image dataset with 348 samples and designed three 3D convolution networks with different structures, and performed a normal/abnormal binary classification experiments on the acquired abdominal CT images. The best performance model in this paper was a 3D convolution networks with three improved residual modules. The average accuracy of normal/abnormal binary classification experiment was 96.2%, and the AUC was 0.989, which was 2.2% and 2.9% higher than the baseline model. The experiment results showed that the 3D convolutional networks with have outstanding performance in the CRC normal/abnormal classification task, it also has good generalization ability and only under CT-level weak-labeled data, it is helpful to clinical application.

Key words: 3D convolution, residual learning, CT image classification, deep learning