The invention belongs to the technical field of image recognition, and discloses a pelvic medical image
automatic segmentation model based on global-local feature optimization. According to the model, an efficient non-local attention mechanism and a double attention mechanism are combined, global anatomical structure understanding and local fracture
feature extraction are collaboratively optimized, and a three-level optimization strategy is adopted to achieve
collaboration of anatomical constraint and
pathological response. Specifically, the efficient non-local attention mechanism can enhance the global
perception ability of the model and help the model to better understand a complex pelvic anatomical structure; and the double attention mechanism is helpful for the model to pay attention to details of a local area, such as changes of a fracture edge and a
lesion area, so that the segmentation precision is improved. Meanwhile, according to the scheme, a mixed
loss function combining
label distribution
perception loss and a surface supervision strategy is provided, the segmentation precision of an
edge region is optimized through a guide network, the probability of missing detection and misjudgment is reduced, and then the robustness of the model in a complex
pathological state is enhanced.