Research On Classification Algorithm Of Skin Lesions Image Based On Deep Learning

Skin cancer is a worldwide serious disease,among which melanoma is the deadliest,and the number of cases is increasing year by year.If it can be detected as soon as possible and treated with excision,it can be cured.The biggest difficulty of melanoma is that the lesions are not similar within the categories but there are great similarities between the categories.For the detection of skin cancer,skin lesion dermoscopy images are generally used.This is a non-invasive imaging technique that obtains high-resolution images for lesions by eliminating reflections on the skin surface,magnifying the lesion area,and enhancing the pigmentation of the lesion area.diagnosis.At present,naked-eye screening is mostly used in clinical diagnosis,which relies on the professional knowledge and clinical experience of doctors,which is not only time-consuming,but also subjective.With the development of technology,the introduction of an objective and reliable computer-aided diagnosis system can effectively improve the efficiency of screening and diagnosis,and is of great significance for the early diagnosis and accurate resection of melanoma.Aiming at the problem of inaccurate segmentation of skin lesion images and poor classification accuracy due to the lack of labeled data in the classification of lesion images,this article is based on deep learning,and conducts in-depth research and exploration of computer-aided diagnosis methods from two aspects of lesion segmentation and recognition.,The main research contents are as follows:(1)In order to improve the segmentation accuracy of the lesion area of the skin lesion dermoscopy image,an improved dense convolution network(Dense Net-BC)skin lesion segmentation algorithm is proposed.First,change the connection between the traditional algorithm layer and the layer,through the dense connection so that all layers can directly access the gradient from the original input signal to the loss function,so that the image feature information can be maximized;secondly,to speed up the model Convergence speed,using small convolution kernels in both the bottleneck layer and the transition layer,not only can reduce the number of model parameters,improve the calculation efficiency of the model,but also solve the problem of a large number of channels output.The performance of the Dense Net-BC algorithm and the VGG-16,Inception-v3 and Res Net-50 algorithms are compared on the ISIC 2018 task 1 skin lesion segmentation data set.The accuracy of lesion segmentation under the Dense Net-BC algorithm is 0.975,and the Threshold Jaccard value is 0.835.The segmentation accuracy is 0.22 higher than Res Net-50 and 0.8 higher than Inception-v3.The experimental results show that the Dense NetBC algorithm can effectively The skin lesion image is segmented.(2)Aiming at the problem of low classification accuracy caused by the lack of labeled data in the classification of skin lesion images,a limited skin lesion image classification method based on meta-learning is proposed.First of all,in view of the problem that the model performance depends heavily on the large-scale labeled data set and the data is not balanced,the double standard that considers the information and representativeness of the sample is adopted for sample selection;secondly,in order to extract the fine-grained features of the image,the The high-resolution image processed by the patch method is used as the input of the multi-scale residual embedding network;finally,the multi-scale embedding network is combined with the meta-learning method based on metric learning,and the samples in the query set are measured by the Euclidean distance function To the distance between the reconstructed features of the embedded network and the class,so as to realize the classification of the skin lesion image.The experimental results on the HAM10000 skin lesion data set show that in the case of limited samples,the model in this paper can quickly adapt to new tasks.The performance test results obtained in 5-Shot 3-Way are increased by 4.16% compared with MAML.Compared with DAML,it is improved by 1.06%,which proves the effectiveness of the multi-scale embedded network method based on meta-learning in the classification of limited lesion data,and has guiding significance for clinical diagnosis.