The extension of fine microscopic cracks in muddy shale during water saturation-deydration circulation has an important role in the propagation of hydraulic fractures and the formation of fracture network. However, traditional image processing methods for segmenting CT scan images of muddy shale are prone to low efficiency and poor accuracy, as well as lack automation and intelligence. This study proposes a muddy shale crack segmentation network (MSCS-Net) based on the U-Net model that fuses the residual network and multi-scale features of Convolutional Neural Networks (CNNs). The proposed MSCS-ett efficiently segmented muddy shale cracks in CT scanned images after a degradation cycle, allowing for both qualitative and quantitative analysis. The results showed that the values of precision (P), recall (R), F1 score (F1_score), Intersection and Union Ratio (IoU) and Pixel Accuracy (PA) of the proposed MSCS-Net were 91.27%, 93.89%, 92.56%, 85.32% and 98.34%, respectively. Besides, the detection performance of the MSCS-Net was also compared with that of the other three different deep learning models (U-Net, U-Net3 + and Attention U-Net). The test results have demonstrated the superiority of the MSCS-Net over the other three network models in crack detection, localization and segmentation.