Segmentation of Metastatic Lesions in Radionuclide Bone Imaging Based on an Improved Attention U-Net Model

The automatic segmentation of metastatic lesions in radionuclide bone imaging holds significant clinical importance. To reliably and accurately segment metastatic lesions in radionuclide bone imaging, this paper presents an enhanced lesion segmentation model based on the Attention U-Net architecture. To augment the feature extraction capability from input images, a Depthwise Separable Convolutional Attention module was constructed. It utilizes depthwise separable convolutions to enhance the extraction of lower-level semantic features from input images and incorporates the CBAM to emphasize attention towards spatial and channel information. Addressing the challenge posed by varying sizes of lesions in radionuclide bone imaging and the emphasis on extracting vital channel feature information, a Multi-scale Convolution of Attention module was developed. It employs multi-scale convolutions combined with a channel attention mechanism to enhance the extraction and preservation of bone scan image features. Experiments on a real clinical dataset of radionuclide bone imaging achieved a Dice Similarity Coefficient (DSC) evaluation metric of 0.7044. The results indicate that the model exhibits excellent performance in segmenting metastatic lesions and holds substantial clinical diagnostic value.