Multi-view feature extraction encoder based on U-Net and linear transformer for adolescent idiopathic scoliosis spine segmentation

Accurate assessment of Adolescent Idiopathic Scoliosis (AIS) is crucial for clinical diagnosis and treatment planning. Compared to X-ray images, Computed Tomography (CT) images provide more detailed information about the spinal structures. However, due to the abnormal curvature of the spine and vertebral torsion in AIS patients, the segmentation of anatomical structures becomes more challenging. Therefore, we propose a vertebral segmentation method based on a multi-view feature extraction encoder and a linear Transformer for the automatic segmentation of AIS vertebrae. First, we designed a multi-view feature extraction encoder. By transforming the input image patches into three different views through dimensionality transformation, these patches are then fed into three independent encoders to extract three-dimensional features structural. Next, convolutional units are used to fuse the multi-view features after dimensionality transformation across channels, capturing the correlation between different views. Finally, the fused multi-view features are input into the linear Transformer module. By leveraging the long-range dependency capture capability of the Transformer, the method captures the relationships between different vertebrae, thereby enhancing the performance of the model. We conducted experiments on a CT image dataset of AIS patients and evaluated the results using the Dice Similarity Coefficient (DSC) and the 95th percentile of the Hausdorff Distance (HD95). The results showed that the DSC values for the cervical, thoracic, and lumbar vertebrae reached 87.79%, 78.93%, and 86.03%, respectively, while the HD95 values were 2.48 mm, 5.39 mm, and 2.59 mm, respectively. Compared to other methods, our proposed method achieved significantly better segmentation results.