Classification Method And Application Research Based On Imbalanced Medical Image Data Sets

Nowadays,as the acquisition technology of medical images rapidly develops,new technology and equipment emerge and develop rapidly.The enormous data of medical images not only enriches the diagnostic method,but also improves the diagnostic accuracy of disease.Getting medical image becomes easier as the medical-image acquisition device prevails.However,the diagnosis requires high-cost trained doctors to interpret the medical images.At the same time,small and medium-sized hospitals lack doctors with such expertise As a result,the medical-image devices are not fully utilized as they designed to be.For application,a medical image can often be used for detecting various diseases.For example,fundus images are able to reflect the medical conditions of diseases such as diabetes,hypertension,glaucoma,and even arteriosclerosis directly.Therefore,the combination of computer vision technology and medical images can be used to screen diseases as early as possible.Discovering the disease in advance affect the prevention and treatment in a positive way.Different from the general classification problem,the data-set in the medical field has a common characteristic – the data set is highly skewed – which the number of normal samples is much higher than the number of diseased samples.Common classifiers trained on such data-sets tend to have obvious “bias”,which means that the diseased samples are incorrectly classified as normal samples.The “biased” sample causing misdiagnosis,which might result the patients miss the best treatment time.In response to the problems above,this article takes the cataract fundus image classification system as the starting point.This paper uses the lightweight web framework Flask built in Python and the neural network framework Caffe to build a cataract remote-classification system based on the 7848 fundus images training residual network combination.At the same time,aiming at the imbalance of medical image samples,this paper introduces some methods to alleviate the problems that caused by imbalanced problems,so as to achieve better classification result and make bring more practical value to the diagnostic system.