Nowadays, the deep learning' development in the image processing field has provided a better method for the segmentation of prostate cancer. Compared with traditional prostate cancer segmentation methods, the prostate can be analyzed by analyzing MRI with the advantages of high tissue resolution and apparent internal structure. Cancer segmentation is not only non-invasive but also has a low rate of misdiagnosis and missed diagnosis. Using deep learning algorithms for MRI analysis does not rely too much on specialized medical knowledge and can save physicians a lot of manual interpretation time. Therefore, research on the application of deep learning in prostate cancer segmentation is of great significance. This project aims to study the prostate cancer segmentation method in multi-parameter MR images based on the deep learning algorithm. Based on the research goals, the following research contents, results, and conclusions have been obtained: Analyze the characteristics and existing problems of the ProstateX data set, and propose image resampling, cropping and normalization, and other pre-processing methods, which not only remove the noise in the MR image but also retain the rich details of the image. Based on the advantages of the U-shaped network architecture, which has the benefits of multi-scale learning features, skip connections to supplement detailed information. Up sampling to restore the feature size, this project adopts the U- shaped network architecture as the basic architecture by setting different modeling methods, i.e., V-Net and U2-Net-based segmentation of single-modality MRI prostate cancer, V-Net-based single-modality MRI prostate and U2-Net-based segmentation of prostate cancer, U2-based -Net's multimodal MRI segmentation of prostate cancer, which can gradually and effectively improve the segmentation effect and accurately realize the prostate cancer' segmentation. In the U2-Net-based multi-modal MRI prostate cancer segmentation method, post- processing methods such as removing small objects at the edge and retaining the maximum connected domain can be used to optimize the U2-Net model and achieve better prostate cancer segmentation results.