The Segmentation of Knee MR Image Using Nested Deep Network and Attention Mechanism

Purpose: Magnetic resonance imaging is of great significance in clinical diagnosis of knee joint. Constructing knee model based on the segmentation of MR images is important for many applications, especially the local specific absorption rate estimation which needs electromagnetic simulation. Methods and Materials: In this paper, we proposed a method using a nested deep network, U-Net++, and attention mechanism to strengthen the segmentation effect of the tissues. The residual module was used to enhance the convergence ability of the network. In addition, Multi-scale deep supervision was performed to preserve the rich semantic features of the decoder paths. Furthermore, we also used a multiclass classification guided module to reduce false positives in image segmentation and improve the overall accuracy of image segmentation. Results and Discussion: Compared with current mainstream methods, the proposed method achieved better performance of tissue segmentation on the dataset of T1-weighted sagittal images collected by our laboratory, especially for cartilage and meniscus. Conclusion: The method integrating U-Net++ and attention mechanism is hopeful to be used to construct knee model for local specific absorption rate estimation.