Crack segmentation in CT images of mud shale is a key link in obtaining crack information, which is of great significance in revealing the spatial distribution of cracks and clarifying the reservoir bed characteristics. Aiming at the problems that traditional image processing methods are unable to segment mud shale cracks autonomously, as well as the low segmentation efficiency due to the complexity and variety of crack morphology in CT images of mud shale, large scale changes, and the similarity of gray scale values between the small cracks and the surrounding rock body, a convolutional neural network MCS-Net, which integrates residual network and multiscale features, is proposed to realize the highly efficient segmentation of mud shale cracks through the improvement of the U-Net network model. The proposed network consists of three parts: feature extraction, special fusion and prediction output. The feature extraction part uses the residual module instead of the conventional convolution to fully obtain the shallow feature information and the high-level semantic information; the feature fusion part adopts the multi-scale convolutional fusion module to obtain the multi-scale feature information of the cracks, and then up-sampling layer-by-layer to restore the resolution of the image and accurately locate the cracks; and the prediction output part ultimately generates a binary mud shale crack segmentation map. The experiments are tested on the constructed mudshale crack dataset MudshaleCrack and compared with six mainstream deep learning semantic segmentation networks. The results show that the proposed MCS-Net outperforms the comparison networks, with an evaluation metric IoU of 85.32% and an F1 value of 92.56%, which improves the IoU and F1 values by 0.042 8 and 0.090 3, respectively, when compared with the pre-improvement U-Net, which proves the feasibility of the proposed method for mudshale crack segmentation.