Among General Spine Segmentation with Multi-scale and DiscriminateFeature Fusion

Automatic spine segmentation from X-ray images is an important step for diagnosing spinal diseases like scoliosis. However, manual segmentation is time-consuming and prone to errors due to subjective judgments. Thesis proposes a supervised convolutional neural network for accurate and efficient spine segmentation based on X-ray images. The proposed network adopts DUCK-Net, a U-Net structure with six parallel convolution paths, as the backbone and introduces several improvements. To detect vertebrae with different sizes, we introduce Attention Gates between encoder-decoder layers to strengthen multi-scale feature fusion. Channel Interaction Attention block is proposed to enhanced feature fusion process for more discriminate feature representation. Additionally, a curvature loss is included as a regularization term during training to discourage connected vertebrae segmentation. We evaluate our method on a spine segmentation dataset and a polyp segmentation dataset, showing that it achieves reliable performance on Dice coefficient, Jaccard similarity, Precision and Recall. Our model have achieved state-of-the-art performance in spine segmentation from X-ray images and has been implemented in an automated scoliosis diagnosis system in hospital, which shows significant clinical application value and theoretical significance.