Skin Lesion Classification Based On Fusing Deep-Learning-Feature And Multi-Feature

With the rapid development of computer technology,computer-aided diagnosis system(CAD)has been applied in medical institutions increasingly.Diseases which are easy to be identified could be diagnosed by CAD accurately.The detection of dermatosis in the initial stage mainly depends on the analysis of dermoscopy images,leading to the increasing demand for intelligent diagnosis by image processing.However,due to the different collection environment,different stage of the disease and various patient’s skin color,there are great differences in images even if they are the same skin disease,which resulting in low accuracy of intelligent diagnosis.In addition,the feature extraction of skin lesion images mainly relies on neural network,making it difficult to extract the specific feature with significant contribution for classification.This paper aims to explore the above technical bottlenecks in skin lesion classification.The main research contents are as follows.(1)This paper proposes an improved lightweight U-Net based on the classical U-Net,aiming at the problem that computer resources of medical institutions are always difficult to meet the high computing requirement of image segmentation in skin lesion classification.The lightweight deep-learning model has been designed by streamlining the structure of the network.Magnitude of parameters has dropped from ten million to one million in lightweight deeplearning model,while the accuracy is retained the same as the classical model.Through the multi-scale fine-tuning method in this paper,the global information and local information on dermoscopy images are merged to improve the segmentation accuracy without further increasing the complexity of the network.As a result,the Jaccard index reaches 0.724 and the dice coefficient reaches 0.814 in the improved lightweight U-Net.(2)Considering the importance of LBP feature for skin lesion,this paper studies multifeature fusion method based on fusing deep-learning features and hand-crafted features.In this paper,the confidence-fusion method and feature-fusion method are proposed.The confidencefusion method improves the accuracy by weighting the confidence of the LBP model and the RGB model,which achieves 98.0% accuracy.In addition,the feature-fusion method merges LBP features and deep-learning features in the feature layer and reach the accuracy rate of97.6%.(3)An IOT(Internet of things)dermoscopy system is designed with the above research method in this paper.Due to the lack of dermoscopy detection system and scheme for computeraided diagnosis with good portability on the market,the intelligent IOT dermoscopy system in this paper has great advantages potentially in diagnosis accuracy and portability.The system consists of handheld dermoscopy,dermoscopy server,online expert system and patient application.It applies the skin lesions classification method based on the deep multi-feature fusion proposed in this paper and realizes the linkage of each part by IOT and SAAS,which makes the diagnosis process of skin diseases conveniently and reliably.