Comparative Evaluation ofclassifiers andFeature Selection Methodsfor MassScreening inDigitized Mammograms
暂无分享,去创建一个
In thispaper, threegroupsofcharacteristics related tomasstexture areadopted, namely, SGLD (Spatial GrayLevelDependence), TS(Texture Spectrum) andTFCM (TextureFeatureCodingMethod)to describethe characteristics ofmassesandnormaltextures ondigitized mammograms.Next,underthetesting byclassifiers, three FeatureSelection Methods--SBS (Sequential Backward Selection), SFS(Sequential ForwardSelection) andSFSM (Sequential Floating SearchMethod)areusedtofindout suboptimal subset from19features inordertoimprove the performance ofmassdetection. Finally, twoclassifiers PNN (Probabilistic NeuralNetwork) andSVM (Support Vector Machine) areapplied andtheir performances arecompared. I. INTRODUCTION The majorsymptomofbreast canceriseither microcalcifications and/ormassesin mammograms. However, up tothismomentmassesdetection from digitized mammogramsisstill verychallenging. Themain reason iscaused bythat themasses usually mixwiththe inhomogeneous tissues inthebreast. Thegraylevels of those inhomogeneous tissues inthebreast could varywith thedistribution ofbreast softtissue. Furthermore, the difficulty could beincreased duetothat themasses shown indigitized mammograms aresimilar totheglands, cysts or dense portions ofthebreast (1). Comparing withusing thehistogram statistic, variation ofgraylevels orthecontrast ofintensities, usingthe variation oftexture features isreportedly morereliable in thedetection ofmasses(1). Basedonthis consideration, oursystem usestextures forthediscrimination ofedge, background andmassblocks fromtheROIs.Firstly, breast region isextracted. Then, Gradient Enhancement and MedianFiltering areperformed forimages enhancement andremoving noises. Next, texture structure ofimages is analyzed using SGLD,TSandTFCM (2,3), and19vectors fordescribing texture features arecalculated fromevery 32*32pixel blockwithin extracted breast region. After feature selection, thesuboptimal feature subset isthen entered into trained classifiers, whichareusedtoidentify if there isanymassinevery block ofbreast region. Finally, thissubsystem utilizes optionally PNN (Probabilistic Neural Network) andSVM (Support Vector Machine) (4) classifiers forComparative Evaluation tooptimize the system performance.
[1] B. Aldrich,et al. Application of spatial grey level dependence methods to digitized mammograms , 1994, Proceedings of the IEEE Southwest Symposium on Image Analysis and Interpretation.
[2] Stanley J. Reeves,et al. Sequential algorithms for observation selection , 1999, IEEE Trans. Signal Process..