Residual Network And Integrated Learning Based Assisted Diagnosis Of Melanoma

As the largest organ of the human body,the skin is susceptible to cancerous changes and melanoma due to its direct exposure to various complex environments,such as germ infections and ultraviolet rays.Currently,the main way to diagnose melanoma is through dermatologists with their related medical knowledge,but since the diagnosis of melanoma needs to rely on professional equipment such as dermatoscope,this method not only consumes a lot of medical resources,but also is prone to misdiagnosis,and the limited medical resources nowadays can hardly meet the diagnostic needs.With the continuous development of computer technology,a new way for melanoma diagnosis has been provided.At present,there are still great difficulties in the process of computer diagnosis of melanoma,because melanoma images have high similarity with other skin diseases in terms of shape,color and other characteristics,which makes the accuracy of computer-aided diagnosis not high.To solve this problem,the main research of this paper is as follows.1.To improve the accuracy of computer-aided diagnosis of melanoma,the attention mechanism is combined with convolutional neural network to design a base classifier with ResNet50 as the backbone network,and the last convolutional layer in the network is replaced with a self-attention module for improvement.After experimental comparison,the improved classifier using the self-attentive module improved the accuracy of melanoma recognition by 0.9% and the AUC(Area Under Curve)by 0.5%,and the attention distribution of the improved classifier was more focused on the lesion site.2.To further improve the accuracy and robustness of the base classifier,we propose to extract the melanoma diagnostic features from the dermatological images,which makes the classifier pay more attention to the effective features that are helpful for diagnosis during the training process.To reduce the useless information in the images and highlight the effective information,the dermatological images are pre-processed.The feature extraction session used texture feature extraction based on GLCM(Gray-level co-occurrence matrix)for melanoma images,color feature extraction based on color diagnostic elements in ABCD rule for melanoma images,and contour feature extraction based on Meanshift algorithm for melanoma images,respectively.The experiments show that after combining the 3 diagnostic features of texture,color and contour,the dermatological classifier improves the recognition accuracy of melanoma by 1% and the AUC by 1.3%.3.A melanoma recognition model based on Adaboost algorithm was designed to integrate multiple base classifiers.The experimental results showed that the performance of the integrated melanoma recognition model was significantly improved compared with that before integration,and the accuracy of melanoma recognition was improved by 2.4% and the AUC was improved by 0.4% when the number of base classifiers was 5.Figure 64 Table 6 Reference 64...