MFRes-UNet: A ResUNet Fusing Multiscale Features for Accurate Spine CT Images Segmentation

Accurate vertebrae segmentation in spine CT images is essential for diagnosing and treating spinal diseases. However, existing methods face challenges due to the extensive span of the spine, the complex shape of its curvature, and the high similarity between adjacent vertebrae in CT sagittal plane images. These factors often lead to limitations in capturing the detailed features and contextual information necessary for accurate segmentation. This paper presents a ResUnet model that incorporates both a multiscale feature extraction module and a feature refinement module. The multiscale feature extraction module utilizes multiscale and dilated convolution techniques to broaden the model's receptive field, thereby enhancing the extraction of global semantic information. Additionally, the feature refinement module employs a decoupling strategy with large convolution kernels and a spatial kernel selection mechanism to enable the model to capture richer, detailed features. The proposed model has been validated on the CTSpine1k dataset, and its performance was rigorously evaluated. The experimental results indicate that the model achieves an average Dice score of 88.89% and an average HD95 distance of 4.17 mm. Compared to other state-of-the-art models, the proposed method exhibits a significant advantage in accurately processing the details of vertebral edges.