To address irregular contours and border ambiguity in skeletal fluorosis X-ray images, A segmentation method that combines dynamic snake convolution and graph convolution attention is proposed. The omni-dimensional dynamic snake convolution module is introduced in the encoder to compensate for the lack of multi-scale feature perception in single convolutions, reducing feature loss. The graph convolution attention module in the decoder captures long-range dependencies and focuses on key lesion areas through channel attention. Additionally,　the ghost convolution block enhances performance and reduces parameters. By eliminating the skip connection with the first two layers of features in the U-Net decoding structure, generalization performance improves. Validation on a skeletal fluorosis X-ray image segmentation dataset shows DSC 0.77, IoU 0.64, Accuracy 0.85, and Recall 0.75. Experimental results demonstrate the method's superiority in segmenting skeletal fluorosis small osteogenic lesions compared to current mainstream medical image segmentation methods.