MachineLearning Techniques inDetecting ofPulmonary Embolisms

MarkH.Myers,IgorBeliaev, King-Ip Lin Abstract - ComputerAidedDetection (CAD)systems haverecently beenusedby physicians to help automatically detect early formsofbreast cancer inX- rayimages, lungnodules inlungCT images, andpolyps incolon CT images. We discuss anautomatic detection mechanism using agenetic algorithms (GA)approach to identify and classify PulmonaryEmbolisms(PE) captured through ComputedTomography Angiography (CTA).Ourmethodenhances theperformance ofthe classification of diseases as comparedto other methodologies discussed inthis paper. I.INTRODUCTION Machinelearning techniques arefastbecoming important toolsinthemedical field (14,15).These techniques havebeenutilized intheareaofdisease diagnostics inordertodetermine thecauseofhealth deficiencies. Fordisease diagnostics, various classification- basedlearning methods likeneural networks, k-nearest neighbors andrule induction arebeing used. Oneproblem that suchmethods faceisthelarge numberoffeatures that aregenerated fromsuchproblems. Forexample, inmany cases, medical images areusedfordiagnostics andeach imagecancontains multiple objects withdifferent features (shape, size, andotherdescriptive features). Thelarge numberoffeatures makesmanysuchmethods impractical. This in term requires us to use feature selection/dimensionality reduction techniques toproduce a small sets ofuseful features fortheclassification methods. Inthis paper, wediscuss several former approaches and provide a newapproach toPulmonary Embolisms (PE) detection. We usea combination ofGenetic Algorithms (GA)forfeature selection andMulti-Layer Perceptron (MLP)inordertoperform classification inorderto accurately discover thepulmonary embolisms. Theproblem isinteresting asithasdifferent performance requirements - suchasthrough misclassification canleadtosignificant undesirable impacts tothepatient. Inthis paper, wediscuss howgenetic algorithms haveimproved theperformance requirements of attribute selection problems. Our experimental results showthat feature selection based onthe genetic algorithm leads tomuchimproved performance for agiven task.