The Research Of Medical Image Analysis Based On Cost Sensitive Boosting

Medical imaging technology plays an important role in medicaldetection. The computer aided medical image analysis, includingclassification, detection and recognition, is expected to raise the detectionrate and serve quantified reference basis to doctors. Furthermore, it iswidely used in review analysis and medical training. The traditionalclassification algorithms, with target to minimize the error rate, treat twotypes of errors (false negative and false positive) equally. However,medical diagnosis is a typical cost-sensitive problem, in which the cost offalse negative decision is rather critical than false positive. Thecost-insensitive models usually perform weak in controlling specific errorrate, and a cost-sensitive extension is needed to solve this kind ofproblem. Our work is focused on the research of medical image analysisbased on cost sensitive boosting.First, we analyze the cost-sensitive model and Boosting algorithm.As simple threshold manipulation is not able to provide prefectcost-sensitive performance in boosing， cost loss function and costweight-update strategy are introduced to change Boosting trainingemphasis to cost-sensitive neighborhood. It is used to derivecost-sensitive extension on AdaBoost, RealBoost and LogitBoost. Theexperiment on UCI dataset is conducted to verify the programmed costalgorithm，and comparing with cost-insensitive Boosting.Second, we discuss the application based on cost-sensitiveBoosting in the recognition between benign and malignant in breastcancer ultrasound image. The image feature is extracted according toBI-RADS (Breast imaging report and data system), and a more simplifiedsub-feature set is obtained through mRMR (minimal redundancy maximalrelevance) algorithm. Three cost-sensitive Boosting models are trainedand optimal classification parameters are set by tests. The experiment shows cost-sensitive AdaBoost performs the best, with AUC (area underROC curve) at0.859in the condition of controlled false negative rate at5%, and three of cost-sensitive model performs better thancost-insensitive ones.And then, we discuss the application based on cost-sensitiveBoosting in the abnormality detection on gastroscopic image. In theimage preprocess, specular reflection detection through IS (Intensity andSaturation) distribution is presented. Considering illuminationinsensitivity and rotation insensitivity, normalized color spaces, rg and hsspace, as well as normalized LBP (Local Binary Pattern) are introducedfor feature extraction. Then, patch method is used for classifier trainingand detection. Three cost-sensitive models are constructed and comparedwith cost-insensitive ones. The experiment shows that cost-sensitivealgorithm performs better in false negative controlling, and detection rateis raised through patch assemble.The cost-sensitive extension of Boosting, with application inmedical image analysis, could control false negative better and performwell generally.