Design And Verification Of Risk Rating Algorithm For Infantile Hemangioma Based On Deep Learning

Infant hemangioma,hereinafter referred to as hemangioma,is one of the most common soft tissue tumors in children.The incidence of hemangioma before 1 year old is as high as 10%-20%,most of which can be cured automatically,but some serious ones can threaten the normal growth and life of the baby.The work distinguishing the risk level of hemangioma accurately and taking corresponding treatment measures is currently mainly done manually by pediatricians with high professional quality.The lack of highly specialized pediatricians and the neglect of the risk of hemangioma by some parents can easily delay the timely treatment of hemangioma.The use of artificial intelligence methods to make an automatic and accurate risk rating for the status of hemangioma is extremely important for the timely treatment of patients.At present,many scholars apply deep learning technology to the diagnosis of hemangioma,distinguishing whether it is hemangioma from the perspective of images,but they do not integrate electronic medical record data,the hemangioma data set they use is very small,and there is no assessment of disease risk,so the clinical utility is low.In response to the above problems,this thesis studies the hemangioma risk rating algorithm based on deep learning and multi-modal feature fusion,and develops a smart phone application.Deploying the model algorithm on the mobile phone facilitates the early detection and treatment of high-risk hemangioma by guardians and primary medical staff.The main research contents and results are as follows:A large hemangioma risk rating data set is constructed.The corresponding patient data is obtained from the Children’s Hospital,including electronic medical record data and image data.For the electronic medical record data,the key information of the patient’s age,the location of the disease and the size of the skin lesion is extracted and then quantified.For the image data,the region image of the hemangioma is cropped and adjusted to the same resolution.For each piece of data,two professional pediatricians perform risk level calibration,and their average value is took as the label for this piece of data.Since the original data set is not large enough and there is a problem of data imbalance,data balancing and data enhancement operations are carried out on the original data set to ensure that each type of data has the same sufficient amount,and the number of final data sets reached 3282.A risk rating algorithm for hemangioma based on multi-modal feature fusion is proposed.For the two kinds of hemangioma data,two neural network models are built to extract data features and merge them.For the structured electronic medical record data,we build a deep neural network model to extract the characteristics of patients with hemangioma in the medical record.For the processed image data,a convolutional neural network model is built to extract the features of the hemangioma patient on the image.After the two models are trained separately to reach the optimal value,the two models are fused.Two different feature information are combined for training.The performance of the fusion model is improved from 88.9% and 90.8% of the single model to 92.6%.A smart phone application is designed,and the model is transplanted to the phone to realize a convenient and quick hemangioma risk rating application.Taking into account the limited computing conditions of mobile phones,for the convolutional neural network part of the fusion model,deep separable convolutions are used to replace the original ordinary convolutions,reducing the amount of parameters and calculations,and improving the running speed of the model.The model is transplanted to a smart phone after quantitative optimization,and the model algorithm is called on the designed mobile phone application to predict the risk level of hemangioma.The experimental results show that under the condition of ensuring that the accuracy rate is not reduced,the task of rating the risk of hemangioma can be completed conveniently and quickly on the mobile terminal.For areas with limited medical resources,it can assist infant guardians and grassroots medical personnel in early identification of hemangioma risk levels and taking corresponding treatment measures.