Research On Dermoscopic Image Classification Method Based On Deep Attention Network

Skin cancer is a common malignant tumor that can cause serious harm to patients if not diagnosed and treated in a timely manner.To improve diagnosis and treatment accuracy,dermatology has proposed skin imaging diagnosis technology,which has become the best method for diagnosing skin lesion types.Traditional skin imaging diagnosis methods require professional doctors to make judgments,but there are various problems that affect the accuracy of diagnosis.In recent years,with the development of artificial intelligence-related technologies,deep learning-based skin imaging classification methods can automatically classify images.However,this method also has some problems,such as imbalanced categories in datasets,intraclass similarity and inter-class differences in lesion images,inadequate utilization of patients’ disease information,and lack of attention to lesion areas in networks.To address these issues,this paper combines a deep attention model to investigate skin imaging classification methods.The main research contents of this paper are as follows:(1)To address the issues of inter-class similarity and intra-class diversity in lesion images,a model based on Grouping of Multi-scale Attention Blocks is proposed.This model utilizes multiple branches of different scale sizes to extract fine-grained features,and enhances the extracted features by capturing the spatial and channel dependencies through grouping.Finally,all sub-features are fused to effectively improve the model’s focus and attention on lesions of different scales.To address the issue of imbalanced sample sizes in the dataset,a class-weighted loss function is used to optimize the model and improve its accuracy.Experimental results show that this method significantly improves several evaluation metrics such as ACC and AUC,and performs well in comparison with other models.(2)In order to address the problem of poor classification performance caused by the insufficient minority class samples in the dataset,this paper adopts appropriate random oversampling and various data augmentation methods to increase high-quality samples of the minority class and improve the generalization performance of the model.Moreover,to address the problem that skin lesion classification models lack the correlation between patient pathological metadata and image data,this paper also utilizes patient metadata and designs encoding forms and branch networks specifically to increase the feature information of lesion types and help the model better understand images.Finally,the attention mechanism of the backbone network is combined with the metadata branch network.Experimental results show that the classification model combining data augmentation and metadata can effectively improve classification accuracy,and also performs well compared to other similar models.