Auto-segmentation method based on deep learning for the knee joint in MR images

Auto-segmentation of the knee joint in magnetic resonance (MR) images is significant for clinical requirements. However, it is challenging due to that the segmentation targets have dramatically different sizes. In this study, an end-to-end DRD U-Net is proposed, which is based on the deep learning framework. The residual module is used as the basic module in the U-Net model, which increases the ability of reusing feature maps. The parallel dilated convolution modules are used to achieve different receptive fields, which can overcome the limitations of single receptive field in the U-Net model and effectively improve the segmentation capability with targets of different sizes. The multi-output fusion deep supervision module is designed to directly utilize the feature maps of different levels. In this way, the information complementarity is obtained, the consistency and accuracy of the segmented regions are improved. The proposed algorithm is evaluated by using the public OAI-ZIB data set. The average segmented surface distance is 0.2 mm, the root mean square surface distance is 0.43 mm, the Hausdorff distance is 5.22 mm, the average dice similarity coefficient (DSC) is 93.05%, and the volume overlap error is 3.86%. Compared with the conventional U-Net and other currently available models, the proposed DRD U-Net has better segmentation accuracy. © 2020, Science Press. All right reserved.