Study On Disease Risk Assessment Based On Multiple Instance Deep Learning

With the development of computer technology and bioinformatics,medical information data is growing rapidly.The medical industry has entered the era of big data.Computer Aided Diagnosis has gradually become a research hot spot of scholars and experts in the industry.As one of the core technologies of computer aided diagnosis,disease risk assessment based on medical images has attracted more and more attention.Aiming at the problems of the current model of disease risk assessment based on medical pathological images,such as using the whole image to learn directly and introducing a lot of noise,which leads to the decline of accuracy,and learning only for the region of interest and ignoring other meaningful parts,which leads to the weak generalization ability of the model,the multiple instance learning method is combined with the deep learning model in this thesis to build a multiple instance learning deep learning model,which is abbreviated as KMIL-DCNN model.The model can highlight the features of the malignant part of the image,enrich the features of the benign part of the image,and assess the disease risk of the medical pathological image more accurately.Firstly,the multiple instance learning algorithm is improved.By combining multiple instance learning with K-means algorithm,a multiple instance learning model based on K-means clustering algorithm,which is abbreviated as KMIL model,is proposed to label sample tags.And the K-means algorithm is improved to determine the initial center point of the class cluster,directly using all negative instances to determine the initial center point of the negative class cluster,which improves the accuracy of KMIL model for instance classification.Then,the multiple instance learning algorithm is combined with a variety of deep learning models,and the input data of the deep learning model is screened.The samples at the edge of cluster in the output of multiple instance learning are discarded,and the samples with more accurate labels are retained as the input of the deep learning model.A variety of deep convolutional neural network models are trained,and the best deep learning model is selected to determine the final KMIL-DCNN model to evaluate the disease risk of medical pathological images.Finally,taking Breakhis breast cancer data set as an example,the effectiveness of the proposed KMIL-DCNN model is verified on the established experimental platform,with accuracy and generalization performance as evaluation indexes.The simulation results show that the proposed KMIL-DCNN model is superior to the existing models in both accuracy and generalization performance,and can carry out more accurate disease risk assessment for medical pathological images.