Improved Transunet's Segmentation Study of Bone Scintigraphy Lesions

In order to solve the problems of irregular targets and fuzzy boundaries in bone scintigraphy segmentation, an improved TransUNet model was proposed. The feature extraction part of the encoder is replaced with an asymmetric convolution residual module to enhance feature capture in different directions and avoid gradient vanishing. At the same time, the cross-fusion module is used to replace the jump connection, which strengthens the deep connection between the encoder and the decoder, suppresses redundant information and improves fine-grained feature capture. In addition, the maximization decision-making method in the two-channel output dimension is used to improve the richness of classification information, capture the uncertain region and reduce the influence of class imbalance in the case of fuzzy boundaries, and obtain a segmentation graph. Experimental results show that the improved Transunet segmentation algorithm has improved the Intersection over Union (IoU), DSC coefficient (Dice Similarity Cofficient), pixel accuracy (CPA) and recall rate (Recall), reaching 0.498, 0.667, 0.662 and 0.761, respectively, which is better than the current mainstream segmentation algorithms. It has certain practical application value.