Research on Automatic Spine Labeling Algorithm Based on IMAA-Unet

Magnetic resonance imaging (MRI) is one of the commonly used inspection methods in clinical imaging diagnosis of human spinal diseases. Rapid and accurate localization and segmentation of vertebrae are helpful to the clinical diagnosis and treatment of spinal diseases. In this paper, according to Nash equilibrium principle, the training data is selectively augmented to obtain more comprehensive training data. Combine traditional Unet network with InceptionV2, Attention Gate, Atrous Spatial Pyramid Pooling to reduce contextual information loss. Then, the Merge module is added to enable the network to learn context features more fully. During marking and positioning, the vertebral body label is set to the coccygeal vertebra (S), the rest of the vertebral body, and the non-vertebral body. Then, the image data is used for training and the recognition model is generated. When it is necessary to mark and locate the spine, the image is recognized according to the trained spine recognition model, and the spatial position information between the recognized vertebrae is calculated. Based on the identification and positioning of coccygeal vertebra, the labeling and positioning of other vertebral bodies are completed. Compared with the results of traditional Unet and Unet++, the algorithm used in this paper has better recognition accuracy.