Melanoma Image Segmentation Based on Improved TransUnet

With the rapid development of deep learning, computer technology and medical imaging technology, medical image analysis has become an indispensable auxiliary means in clinical medicine and medical research. The use of convolutional neural network for medical image analysis has become a research hotspot. As a highly fatal skin cancer, the number of patients with melanoma is increasing year by year. Due to its rapid disease progression and high metastatic tendency, early detection, early diagnosis and rapid treatment are the most effective methods to reduce the risk and mortality of melanoma patients. In view of the fact that melanoma is insipid, difficult to detect and similar to other skin diseases in the early stage, and the limited medical resources and the slow growth of the number of melanoma diagnosis experts, there is an urgent need to develop a painless, simple, rapid and high-precision medical auxiliary model to assist melanoma diagnosis. In this paper, a melanoma image segmentation method based on TransUnet is designed to assist clinical screening, diagnosis, grading, treatment decision-making and efficacy evaluation of melanoma. The ISIC dataset, which is widely used in the diagnosis and research of skin diseases, was selected as the data set. Based on the study of TransUnet network model, an improsved AS-AttTrasUnet algorithm for melanoma segmentation is proposed, which introduces ASPP module and ECAnet channel attention module to retain more feature details and suppress redundant feature information. The experimental results show that the dice of AS-Atttrasunet model on the test set reaches 97.1%, which is verified by comparative experiments that the improved model is significantly better than other classical networks and other transformer models. The effectiveness of ASPP module and ECAnet channel attention module is verified by ablation experiments. After model prediction, post-processing hole filling is performed on the prediction results to further improve the segmentation accuracy of the model. After a large number of tests, the model can meet the speed and accuracy requirements of melanoma segmentation in most scenes and environments.