Research On Classification Algorithm Of Melanoma Based On Dermoscopic Images

Melanoma is a typical and serious skin disease with high mortality rate of melanoma.Currently,the rapid development of deep learning in medical imaging has promoted the improvement of efficiency and accuracy of medical aid diagnosis.Automatic melanoma classification technology based on computer vision technology has a promising future as an effective early diagnostic screening tool that can help physicians quickly screen out suspicious lesions and help improve the efficiency of biopsy or surgery.The development of various new technologies and the growth of data sets have supported the rapid development of various classification algorithms in recent years to reach the judgment level of professional physicians on several medical tasks.However,deep learning in medicine suffers from the difficulties of limited training datasets and imbalanced data,which have been the focus and difficulty of research.Along with the progress of medical deep learning work,research to improve the imbalanced dataset scenario has mainly focused on improving the weak position of minority classes,and image preprocessing and attention techniques have become important research directions to improve melanoma classification algorithms.In this paper,we focus the research on improving impact of imbalanced data and improving model attention to local information direction.We designed two classification models,the main work is as follows:1.Model Ⅰ focuses on improving the difficulty of training melanoma classification models with imbalanced dataset,and improving the comprehensive performance of the models at both data level and algorithm level.First,to address the weak position of minority classes in imbalanced dataset and the training difficulties caused by imbalanced dataset,this paper designs an improved random oversampling and generative adversarial nets to add samples for the malignant class.The enhancement of minority class’ s samples can accelerate the convergence of training and improve the role of minority classes in the backward propagation of updated weights.Second,to address the lack of high-quality samples,the model employs various data enhancement methods,including geometric transformation and noise injection,to improve the general performance of the model.Third,the model also incorporates patient metadata training and a neural network is designed for metadata to supplement the feature information of minority classes.Finally,to improve the network’s attention on important channels in multiple channels,the model uses autonomously learned channel attention and weights the multiple channels.We conducted experiments using these methods step by step,and the results show that the classification performance of the model is gradually improved,and it also performs well in comparison with other models,and the model can effectively improve the performance of melanoma classification models.2.Model Ⅱ mainly focuses on improving the model’s attention to important information based on the attention mechanism.The model applys attention mechanism in several dimensions to improve the ability of the model to learn the important knowledge.First,to improve the model accuracy and put more effort on learning the important focal regions,the model learns the attention distribution in the spatial dimension and weights the feature maps for the low-level feature phase in the convolutional network to improve the model’s interest in the lesion regions.Second,to achieve inter-sample feature enhancement,the model applys the self-attention mechanism to construct a sequential network,and this network learns attention between multiple sample features simultaneously and weights the features of interest.Experiments show that the spatial attention pays attention to the lesion area,and the fusion of spatial attention and selfattention mechanism effectively improves the prediction level of the model.