A generative adversarial network incorporating dual attention mechanisms for the segmentation of knee bone and cartilage.

objectiveThe accurate segmentation of bone and cartilage is important for knee joint assessment. However, the complex shapes of both tissues make segmentation challenging. In this study, a conditional generative adversarial network incorporating dual attention mechanisms is proposed to enhance the segmentation accuracy of bone and cartilage. MATERIALS AND

methodsThe generator (a U-Net) integrates coupling attention mechanisms: squeeze-and-excitation in both the downsampling and upsampling paths, and the attention gate in skip connections, to focus on small tissue regions effectively. The discriminator (a convolutional neural network) fuses feature information from different layers of the generator. The proposed method was trained and tested using images from the Osteoarthritis Initiative (OAI) database. The training set and testing set consisted of 70 and 30 subjects, respectively.

resultsThe experimental results demonstrated that the proposed method achieved satisfactory segmentation performance for four types of tissues: the femur, tibia, femoral cartilage, and tibial cartilage. The Dice similarity coefficient (DSC) values of these tissues were 97.4%, 97.2%, 91.2%, and 88.9%, respectively. Compared with using only GAN, the DSC of femoral cartilage and tibial cartilage increased by 6.9 and 1.7%, respectively. DISCUSSION: By integrating squeeze-and-excitation and attention gate, the network achieved a notable improvement in segmentation accuracy for knee cartilage.