Research On Skin Disease Detection Algorithm Based On Lightweight Network

The incidence of skin diseases is showing a yearly increase worldwide,among which melanocytic lesions may lead to melanoma,which has a high lethality,and early detection and treatment will prolong the patient’s chances of survival.The diagnosis of melanoma mainly relies on dermoscopic images,which are subjective for direct identification by dermatologists,while computer-aided diagnostic systems can effectively improve diagnostic efficiency.Although deep learning techniques have achieved good results in skin disease image segmentation and classification,traditional deep convolutional neural networks tend to have more complex network structures,consume more resources,and are not suitable for deployment on lightweight devices.However,the existing lightweight networks directly used for skin disease detection have poor results.Therefore,reliable lightweight networks are designed to enable them to provide support for computer-aided diagnosis systems.This thesis focuses on the optimization using lightweight networks for skin disease image segmentation and classification,and the main work is as follows:(1)To further reduce the parameters of the skin disease image segmentation network,a skin disease image segmentation algorithm based on Unet and MLP networks is proposed.Firstly,the algorithm utilizes depth separable convolution as the first three layers of the encoder and decoder for reducing the number of model parameter while adding dilation rate to the depth convolution,and thus locating the lesion region more accurately.Secondly,the attention axial shift multilayer perceptron module is proposed in the last two layers of the encoder and decoder,which is beneficial to realize the interaction of different location information,and at the same time strengthen the features of the network in channel and space expression.Finally,experiments are conducted on the dataset ISIC 2018,and the results show that the Dice coefficient of 90.9% is obtained with only 0.98 M parameters.The experiment proves that the algorithm obtains better segmentation results with fewer parameters.(2)To further improve the feature extraction capability of lightweight networks for skin disease image classification,an attention mechanism-based algorithm for skin disease image classification in Mobilenet networks is proposed.The algorithm uses Mobilenet V3 as the backbone network.Firstly,the se SK attention module is proposed to solve the problem that the SE attention module has insufficient ability to capture contextual information and improve the model’s ability to express the features of lesion regions.Secondly,RBN normalization is introduced to enhance the degree of difference between samples,which in turn improves the classification effect of the network model.Finally,by conducting experiments on the HAM10000 and PH2 datasets,and the classification accuracy is 85% and 82.5%,respectively,which illustrates the effectiveness of the algorithm in skin disease classification.(3)In order to take full advantage of the hidden information in the skin disease image segmentation algorithm,a segmentation and classification algorithm that fuses the improved Unet and Mobilenet networks is proposed.The proposed algorithm first performs image segmentation by adding a multiscale module to the segmentation algorithm to retain more texture information of the lesion region.Secondly,the algorithm proposed in third chapter is used for image classification,where the input of the classification is made by fusing the segmented image with the original input image,and the output of the multiscale module in the segmentation algorithm is used as the auxiliary information for classification.Finally,the image segmentation experiments are conducted on the datasets ISIC 2018 and PH2,and the segmentation accuracies of 94.9% and 95.8% are obtained,respectively.The classification verification experiments on the ISIC 2018 dataset obtains a classification accuracy of 87.7%,which is 2.5% better than the image classification task under a single task.The experimental results show that the algorithm assisted by segmentation can effectively perform image classification.