Segmentation of metastatic lesions on bone scan images based on improved UNet3+ network

Automatic segmentation of bone metastases in low resolution SPECT bone scan images has an extensive clinical application. In this paper, we propose an improved UNet3+ network incorporating an attention mechanism to realize automatic segmentation of bone metastatic lesions. Firstly, by improving the full-scale deep supervision module to reduce the semantic gap between low-level and high-level features during feature fusion, which helps to obtain more refined segmentation results. Secondly, the segmentation performance of the model is further improved by adding an attention module to the network. Finally, the paper uses a group of clinical SPECT images to evaluate the custom segmentation model on a test sample of the original dataset with values of 0.6433, 0.6120, and 0.6833 for DSC, CPA, and Recall, respectively. The effectiveness of the model proposed in this paper has been demonstrated through experimental comparisons with several similar models.