IMAGE SEGMENTATION ALGORITHM FOR SKIN LESIONS BASED ON DUAL ATTENTION MECHANISM

In view of the difficulty of melanoma segmentation and the poor segmentation effect in the presence of hair covering, a neural network for skin lesion image segmentation based on dual attention mechanism is proposed, which has two decoding paths and one encoding path. The image was preprocessed and dataenhanced, and the Resnet50 backbone extraction network was used to obtain the feature layer of different resolution sizes. The last feature layer extracted was sampled and fused with the previously extracted feature layer through the first coding path, and the next coding and decoding path was entered. The final output was obtained through RAB space and channel attention module. Comparison and ablation experiments were performed on ISBI2016 skin lesion image dataset for several times. According to the experimental results, excellent segmentation results were obtained for images blocked by hair or other objects. The indexes of the experiment are as follows: accuracy 96.19%, sensitivity 93.32%, specificity 97.32%, Dice coefficient 93.26% and Jaccard coefficient 87.36%, which are all superior to the existing algorithms.