Automatic Recognition And Application Of Skin Lesions Based On Deep Convolutional Networks

Skin disease is one of the most common human diseases,and malignant skin lesions represented by melanoma have a high mortality rate.The key to melanoma treatment is early detection and treatment,but early melanoma is easily confused with melanocyte nevus and seborrheic keratosis in benign skin lesions,resulting in treatment delay.At present,the diagnosis of clinical skin diseases mainly depends on the doctor’s visual observation with clinical diagnosis experience,and lacks scientific quantitative methods.Computer aided diagnosis system for skin lesions can help doctors analyze and judge the condition,improve doctors’ diagnosis efficiency and reduce workload.To solve the problem of insufficient training data of dermoscope images,similarity between classes,intra-class variation,and interference with useless background information,this paper studied the synthesis of dermoscope images and recognition of skin lesions from two aspects in order to automatically identify skin lesions.The specific work is as follows:(1)Aiming at the problem of insufficient training data,this paper proposes a dermoscopy image synthesis algorithm based on generative adversarial network for data augmentation.The algorithm combines the shallow and deep features of the generator to avoid losing the detailed information of the image,and introduces the standard deviation feature matching loss to modify the feature matching loss slightly to stabilize the training of generative adversarial network.Finally,we established a meaningful knowledge model of skin lesion to generate high quality,high resolution dermoscopy images.Compared with DCGAN and PGAN,the dermoscopy images synthesized by the algorithm not only are of higher quality,but also can provide additional information gain for supervised learning to improve the performance of skin lesion classification.(2)Aiming at the problem of background information interference,this paper proposes a two-stage skin lesion recognition framework based on deep convolutional networks.The framework enables the classification network to extract more representative and specific features based on the segmentation results rather than thewhole dermoscope image.For lesion segmentation,this paper first introduces the idea of transfer learning to improve the encoder part of U-Net,and then uses the Dice coefficient loss function to give more attention to the lesion area.For the classification of lesions,we first use the color constancy algorithm to process the dermoscopy images to avoid the problem of color space shift under different light source standards,and then we design a Focal Loss function that incorporates a sample balance factor for the problem of class imbalance.The method in this paper achieved an optimal Jaccard Index segmentation performance index of 0.768 and the optimal average average AUC classification performance index of 0.920 on the ISIC-2017 dataset.(3)Based on the solution of the above problems,using the trained segmentation and classification model,we established an intelligent diagnosis system for skin lesions.Under the framework of two-stage skin lesion recognition,the system can obtain segmentation and classification results by uploading the dermoscope images.This will help accelerate the intelligentization of dermatological diagnosis,alleviate the shortage of medical resources and the problem of imbalance,and bring great convenience to the early screening and accurate diagnosis and treatment of patients.