Research On Content-based Dermoscopy Image Retrieval Method

Skin cancer and various pigmented skin diseases are seriously threatening human health.At present,dermatologists mainly diagnose pigmented skin diseases by observing and analyzing the characteristics of lesions in dermoscopic images.Due to the small differences between different types of skin disease lesions,it is difficult for doctors to judge the type of lesions through naked eye observation.On the other hand,with the development of big data technology,medical data has exploded,and a large amount of dermoscopy image data has overwhelmed dermatologists.Therefore,it is of great significance to realize the efficient retrieval of similar images using the case data in the already diagnosed skin disease database.This thesis takes the data of seven types of skin diseases in the ISIC2018 dermoscopy dataset as the research object.The main research contents are as follows:In view of the small number of samples in the data set and the uneven distribution of samples in the data set,data enhancement of dermoscopy images,expanding and balancing the data set is beneficial to the training of subsequent models.Aiming at the problems of a wide variety of skin diseases,small differences between categories,and large differences in features within categories,a dermoscopic image retrieval method based on convolutional neural network and hashing is proposed.The advantage is that the two are placed in the same network structure to learn,so that the feature representation of the image and the hash code can be learned simultaneously in the convolutional neural network,and the fast retrieval of similar images in the database image can be realized by the Hamming distance metric method.The experimental results show that the retrieval model based on Res Net50 network and64-bit hash code has better retrieval effect,the average precision is 72.52%,and the retrieval time is4.436 s.In order to further improve the performance of the dermoscopic image retrieval model,combined with the fact that doctors and patients pay more attention to the first few images in the retrieval results,this thesis proposes a re-ranking method based on neighbor propagation.Determine a candidate set with a short Hamming distance from the query image through hash coding,take the high-dimensional features in the candidate set as data points in the neighbor propagation cluster,and find the location where the query image is located through information transfer between data points.Clusters are sorted from small to large according to the principle of distance measurement.The experimental results show that this method can effectively optimize the retrieval ranking,and the average accuracy of the first 10 returned images of the re-ranking method based on neighbor propagation is 2.46% higher than that of the Hamming distance metric method.