Research On Dermoscopy Image Classification Method Based On The Ensemble Of Individual Advantage And Group Decision

Skin cancer is one of the most common cancers in recent years,and it affects all ages.Early diagnosis is very important for the treatment of skin cancer.If it is found early and treated in time,its efficacy and prognosis are good.At present,the classification technology based on dermoscopy images is an effective way to solve these problems.However,manual classification is highly dependent on the clinical experience of doctors,and the dermoscopy image itself is relatively complex.Problems such as large intra-class differences,high inter-class similarity,and blurred skin lesion boundaries pose huge challenges to classification.In addition,due to the differences in technical level and clinical treatment experience,doctors have limited human subjective judgment.Therefore,doctors using computer technology to diagnose and treat patients have gradually become an important research direction in the development of the medical field.The automatic classification of dermoscopic images can not only solve the problem of pathological reading,but also improve people’s awareness of self-protection.At present,the automatic classification of dermoscopy images still has a lot of room for improvement in evaluation criteria such as accuracy,sensitivity,and specificity.Aiming at the problem that the effect of automatic dermoscopy image classification is difficult to meet clinical needs,this paper proposes a dermoscopy image classification method based on the ensemble of individual advantage and group decision.The method includes the ensemble strategy of maximizing individual advantage,the ensemble strategy of maximizing individual and group advantage,the ensemble strategy of block-integrated probability,the ensemble strategy of block-integrated voting.Firstly,the image is preprocessed,such as using image normalization,image denoising and other technologies.Then data augmentation techniques such as generative adversarial networks and image rotation are used to enrich and balance various samples.Then the weighted cross entropy loss function is used to alleviate the problem of uneven distribution in the dataset.Finally,through the method of transfer learning,the pre-training convolutional neural network models are used for fine-tuning,and the effects of different convolutional neural network models on different categories of dermoscopy image classification are compared.The better convolutional neural network models are selected to different ensemble strategies,so as to achieve the classification of dermoscopy images.The method in this paper is performed on the dataset provided by the International Skin Imaging Collaboration(ISIC).Different ensemble strategies have different advantages.Among them,the comprehensive performance of the ensemble strategy of block-integrated voting is better,the classification accuracy is 0.934,the area under the receiver operating characteristic(ROC)curve is 0.931,the sensitivity is 0.695,and the specificity is 0.952.Compared with the single network model,this strategy can better improve the generalization ability of problem solving.Compared with the traditional ensemble strategy,this strategy has stronger robustness and can improve the evaluation criteria to some extent.The method proposed in this paper can help dermatologists to make better diagnosis and treatment to a certain extent.