Research On Skin Cancer Diagnosis Method Based On CycleGAN And Ensemble Learning

Malignant tumors on the skin are known as skin cancers.Currently the best way to diagnose skin cancer in major medical institutions is by using dermoscopic imaging techniques.However,the physical characteristics of this type of disease are not obvious in the early stage of development,and therefore,diagnosis by dermoscopy is more difficult.The complexity of the skin disease images themselves and the own defects of standard image datasets make the deep convolutional neural network model-based skin cancer diagnosis method extremely challenging.In this paper,we focus on the impact of data enhancement on skin cancer diagnosis,optimize the generalization ability of different deep neural network models by integrating them,and creatively propose a patching strategy to improve the effective semantic focusing ability of neural network models,so as to improve the skin cancer diagnosis.The contributions of this paper are mainly in the following three aspects:(1)The dataset in this field is characterized by an insufficient number of samples and an uneven distribution of the number of images belonging to different categories.It is very easy to have classification bias,which affects the accuracy of diagnosis.Therefore,this paper proposes an improved Cycle GAN-based skin cancer diagnosis method.The method introduces attention mechanism and adaptive residual block structure in the generator,and improves the loss function.It enables the improved Cycle GAN-based skin cancer diagnosis model to extract skin lesion features more effectively during the image transformation process.The experimental results show that the method can improve Cycle GAN and enhance the generation of skin diseases to improve the diagnostic accuracy.(2)Although the conditional image synthesis technique mentioned in(1)above can balance the dataset and alleviate the problem of imbalanced number of different types of skin lesion images.However,the generalization ability of a single deep convolutional neural network is weak.Therefore,this article proposes a skin cancer diagnosis method based on semantic enhancement and ensemble learning.This diagnostic method has three stages of operation.Firstly,the improved Cycle GAN is applied offline to synthesize underrepresented category samples from highly representative category samples,completing the conditional image synthesis task;Secondly,we use different image preprocessing methods during the training process;Finally,the trained images are fed into the BAFCNN classification network and the classification results are ultimately obtained.The experimental results indicate that this method can effectively improve the generalization ability of the classification network and further enhance the diagnostic effectiveness of skin cancer.(3)Although(2)above can effectively improve skin cancer diagnosis by enhancing the generalization ability of the classification network.However,the method lacks effective semantic focusing ability on lesion regions.Therefore,this paper proposes a skin cancer diagnosis method based on soft attention mechanism and patching strategy.The diagnosis method consists of three stages: firstly,we synthesize target image samples using Cycle GAN technique to balance the dataset;secondly,we propose a pioneering patching strategy to obtain patch images of the training set;finally,the obtained images are fed into an online feature extraction network to obtain skin lesion image classification features and output classification results.The experimental results show that the method can accurately focus the effective semantic information of lesion regions and improve the diagnostic accuracy of skin lesion images.