Research on Aortic Segmentation Algorithm Based on Capsule Networks

CoA is a common congenital heart defect that can lead to severe consequences if not treated promptly. Clinical diagnosis primarily relies on observations of CT and manual measurements of the aortic diameter, which can be affected by factors such as image quality and the prior knowledge of physician, potentially resulting in missed or misdiagnosed cases. Recent advancements in artificial intelligence technologies have spurred a significant surge in research within the field of smart healthcare, utilizing deep learning and other techniques for disease diagnosis and lesion segmentation. CoA can lead to varying degrees of aortic deformation, making CT rich in morphological information. But existing CNNs are insufficiently sensitive to morphological changes and fail to adequately extract these features, thereby providing unreliable medical references for clinicians. Capsule network, as a specialized neural network model, effectively captures and transmits the morphological characteristics of targets through vectors. However, high input resolutions may lead to parameter explosion and longer training times. Therefore, this paper proposes DEUN that combines a capsule encoder and a convolutional encoder. In this model, the capsule encoder is responsible for extracting the morphological features from CT images, while the convolutional encoder extracts spatial features, thereby leveraging the complementary strengths of both encoders. Additionally, this study innovatively designs CAG and MSLF based on the characteristics of capsule networks and the deep supervision mechanism. The CAG is used to enhance the model's focus on key features while suppressing noise and interference. The MSLF helps to avoid the problem of gradient vanishing, improving the model's robustness and convergence speed. Experiments on a CoA dataset have validated the model's effectiveness in segmenting aortic CT images.