Deep Learning Based Multi-classification Algorithm For Skin Lesions

In recent years,the incidence of skin cancer has continued to increase rapidly worldwide,and once it develops it is usually difficult to cure,so its early diagnosis and treatment are crucial.However,the analysis of skin lesion images is challenging due to high inter-class similarity and intra-class variance.Automated classification of dermatological images using computer technology can help physicians make faster judgments about the condition and improve diagnostic efficiency.In addition,automatic classification techniques based on skin lesion images can also play an important role in areas where there are not enough doctors to reduce patient mortality.For the problems of hair interference in skin lesion images,this paper uses operations such as hair removal to preprocess the dataset.Three convolutional neural network models are also improved,aiming to improve the ability of convolutional neural networks to classify multiple skin lesions,while reducing the complexity and computational effort of the models.The main research is as follows.1.In this paper,we improve a skin lesion classification model based on visual geometry group network(VGG-16)fused with residual network.Firstly,this paper removes hair noise by improving the hair removal algorithm,while applying six data augmentation operations to balance the number of rare diseases.Secondly,VGG-16 is improved and feature loss is reduced by adding a preprocessing layer(CBRM)and fused residual network.Finally,the model was evaluated on the ISIC2018 dataset and experiments showed a test accuracy of 88.14% with a macroaverage of 98%.2.In this paper,a two-stream network-based model for skin lesion classification is improved.Firstly,this paper uses a densely connected network(Dense Net-121)and an improved VGG-16 to construct a two-stream network to compensate for the shortcomings of a single network.Second,in the feature fusion module,we construct multiple receptive fields to obtain multi-scale pathology information and use generalized mean pooling(Ge M-pooling)to reduce the spatial dimensionality of lesion features.Finally,a skin lesion classification system is developed using the improved two-stream network model in this paper.The experimental results show that the accuracy of the model proposed in this paper was 91.24% on the ISIC2018 dataset for testing,with a macroaverage of 95%.3.In this paper,we improve a lightweight network-based classification model for skin lesions.Firstly,Mobile Net-V2 with low model size and computational complexity is used as the backbone network.Secondly,this paper incorporates an improved coordinate attention mechanism(RICA-block)in the inverted residual block to perform different feature extraction for lesioned and non-lesioned regions of skin lesion images.Finally,the model designed in this paper has only 3.32 M model parameters,and the test accuracy is 89.52% and 93.24% on the test datasets of ISIC2018 and ISIC2019,respectively,with a macroaverage of 95% for both.