| Malignant melanoma is a kind of skin cancer with a high mortality rate.Because of the diversity of melanoma types and complex lesion characteristics,it increases the difficulty of clinical diagnosis.Early diagnosis and clinical intervention can greatly reduce the mortality of patients.Dermoscopy technology can obtain clear and enlarged lesion characteristic information,which is helpful to help doctors analyze and diagnose the disease.If only doctors diagnose dermatoscope images with naked eyes,there will be problems of high cost and low accuracy because the number of doctors with rich clinical experience is small,and doctors’ personal subjective factors will affect the diagnosis results.Nowadays,the most efficient and cost-effective solution is to use computer technology to automatically assist the diagnosis of dermoscopic images(mainly including the classification and segmentation of dermoscopic images).However,due to the problems of uneven color,blurred boundary and foreign body occlusion of dermoscopic images,it is more difficult to automatically classify and segment dermoscopic images.In recent years,with the continuous development of computer technology,deep training of data using deep learning can effectively obtain the deep features of images,which is outstanding in the field of medical image classification and segmentation.Therefore,based on the existing algorithms,this paper systematically analyzes and studies the dermoscopic images,and proposes an automatic classification and segmentation algorithm of dermoscopic images based on deep learning.The main results of this paper are as follows:1.An improved network model based on pit(pyramid pooling transformer)is proposed to automatically classify the dermoscopic images of seven types of skin lesions.In this paper,the model is mainly composed of pit module and anti-interference module.Pit inherits the advantages of Vit and improves the robustness of the model by pooling space size conversion.The pre trained pit network has a large number of natural image features,and some pit networks can provide the required image features for downstream classification tasks.In this paper,an anti-interference module is designed,It is used to resist the influence of interference factors(such as hair and foreign body occlusion)in dermoscopy images,so as to improve the model performance and classification accuracy.The experimental results show that the classification accuracy,precision,recall and FL score of this model on ISIC 2018 verification set are as high as 91.58%,83.59%,89.92% and 86.34% respectively,and the number of frames per second(FPS)reaches 85 Hz.Compared with several existing advanced classification networks,the classification performance and model efficiency are improved,which has comparative advantages,which proves that this model has certain practical value.2.An automatic segmentation algorithm of dermoscopy image is proposed.Firstly,resnet-34 is used to extract multiple resolution features,and the transformer module is used to globally model the input features in the context part;Then,the multi-scale information of context features is aggregated through the hybrid pooling module,and an efficient convolution module is designed between the jump connections corresponding to the codec to improve the edge thinning and anti-interference ability of the jump path;Finally,the decoder is used to restore the image resolution and fuse other shallow resolution features layer by layer.Experiments are carried out on the public data sets of ISIC 2017 and ISIC 2018.The focal loss function is used to improve the accuracy of difficult segmentation targets,and compared with other skin segmentation models.The results show the effectiveness of the algorithm.Several groups of experiments show that the automatic classification and segmentation algorithm of dermoscopy image proposed in this paper obtains higher accuracy,and has a certain reference value for the research of computer-aided clinical diagnosis. |
PDF Full Text Request