Research On Online Skin Disease Diagnosis And Recognition Based On Deep Learning

Skin is the first line for preventing human body from exterior substances.However,due to its direct exposure to the air and sunlight,it is easy to get infected with skin diseases due to the environment,food,heredity and physical fitness.Skin cancer is one of the five most lethal cancers in the world.Studies have shown that early diagnosis and timely treatment are one of the most effective ways to reduce the fatality rate.However,the relative shortage of medical facilities resources cannot meet patients’ requirement.Computer aided diagnosis systems based on artificial intelligence technology can not only greatly alleviate the dilemma of insufficient medical resources,but also improve the efficiency of skin disease diagnosis,and provide objective analysis results for doctors.Nevertheless,existing studies seldom take skin disease recognition models’ interpretability into consideration,and the performance are far from satisfaction on large scale datasets.Therefore,exploration towards high-performance system and systematic discussion about model interpretability are of great significance.To improve the performance and interpretability of skin disease recognition model.We propose two novel deep learning models named CFPNet and ANGELE algorithms.Additionally,we develop an online medical image analysis platform named MCloud,to help other researchers have easy access to medical AI abilities.Finally,a knowledge base of dermatology based on the unstructured and semi-structured data from the Internet is constructed,to meet the diverse needs of medical AI in practical application scenarios.The main contributions of this paper are as follows.(1)We propose a novel deep model named Cascaded Feature Pyramid Network(CFPNet),to recognize skin disease from diverse human organ.Experimental results indicate that our proposed CFPNet achieves state-of-the-art performance on the challenging benchmark SD-198 dataset with 68.72% accuracy and 66.15% sensitivity.CFPNet also achieves best results on ISIC datasets,with 92.83% accuracy,53.97% sensitivity and 54.62% F1-score on skin cancer categorization task,and 95.37% accuracy and 74.91% F1-score on benign/malignant recognition task.All of them are the best results at present,compared with other methods.(2)Since the appearances of many skin diseases have modest diversity among different classes,and great diversity within the same class.We apply metric learning to skin disease recognition problem,to improve the feature learning ability of the model and enhance the classification accuracy.Extensive experiments demonstrate the effectiveness of the proposed method.(3)Different from natural image recognition task,medical AI system requires satisfactory interpretability of the model,but the majority of mainstream research works in skin lesion analysis do not consider models’ interpretability.In this paper,Grad-CAM is adopted to generate heatmap of the severely affected area,and the inner characteristics of feature learning is also systematically analyzed.On the one hand,the interpretability of the model can be provided from the perspective of visualization.On the other hand,it can provide basis for doctor’s reexamination.(4)Aiming at the requirement of medical AI for highperformance models,we propose a novel deep learning ensemble model named ANGELE(Adaptive uNcertainty Guided dEep Learning Ensemble).By modeling the output confidence of the base learners,the computational complexity of the ensemble model can be heavily reduced.Experimental results indicate that ANGELE outperforms other methods and achieves state-of-the-art performance on SD-198 with 67.41% accuracy and 66.48% sensitivity.ANGELE also achieves best results on ISIC datasets,with 92.79% accuracy,53.39% sensitivity,and 53.67% F1-score on skin cancer categorization task,and 95.22% accuracy and 74.79% F1-score on benign/malignant recognition task.All of them are the best results at present,compared with other models.(5)In order to combine the research achievements with the practical application environment,an online skin lesion analysis cloud platform named MCloud based on Python is introduced in this paper.It provides dermatological disease recognition service in form of RESTful APIs.In addition,a knowledge base of dermatosis is constructed to provide information about the prevention and treatment of dermatosis,diagnosis and major hospitals.The knowledge base can provide better medical AI services for users.