20050831#lab conference#김진성
TRANSCRIPT
Computer-Aided Diagnosis System for Ground-Glass Opacity
using MDCT Images
2005. 8. 31 Jin Sung Kim
2005 RDMI Lab Conference
Contents
• Introduction– Ground Glass Opacity– Purpose & Idea
• Methods– Concept of Algorithm– Image Processing Module– Texture Analysis– Support Vector Machine
• Further Study
2005 RDMI Lab Conference
Ground Glass OpacityIntroduction
2005 RDMI Lab Conference
Introduction
• Focal ground-glass opacity (GGO) is a finding of early adenocarcinoma or its precursor
• Ground-glass opacity (GGO) detection becomes more simple & efficient after extraction of vessels & solid nodules using 3DMM algorithm
• In this exhibition, I will describe our automated GGO nodule detection program that takes advantages of 3D volumetric data from multi-slice CT
• Japan-Korea Joint Symposium on Medical ImagingJapan-Korea Joint Symposium on Medical Imaging
2005. 9.21~22. 서울 고려대학교 구로병원
Introduction
2005 RDMI Lab Conference
Purpose & Idea
• Previous research groups– General 2D slice CT image– Neural Networks (MLP)
• This Study– 3DMM algorithm – GGO Enhanced Image– Support Vector Machine
Introduction
2005 RDMI Lab Conference
Material
• 10 patients have GGO nodule
• 120KVp, 120 effective mAs
• 3.2 mm slice thickness
• Average 126.9 images/patient
• Programming based on Matlab
• OSU LIBSVM in matlab
Materials & Methods
2005 RDMI Lab Conference
Methods
Air Component
Soft TissuePulmonary VesselSolid nodules
GGO nodules
CT Noises
After soft tissue & air component extraction, GGO detection is more easier !!!!.
IntroductionMaterials & Methods
2005 RDMI Lab Conference
Overall AlgorithmIntroductionMaterials & Methods
2005 RDMI Lab Conference
3D Volume of segmented lung regionMethodsMaterials & Methods
2005 RDMI Lab Conference
3D Image of Pulmonary Vessel extraction using 3DMM algorithm
MethodsMaterials & Methods
The GGO was not include in vessel
We can find a GGO in right lung region
Original CT Image – Soft Tissue Image Using thresholding, GGO can be found
2005 RDMI Lab Conference
ROI matrix, texture analysis
• 32x32 matrix
• Texture– Mean– Standard
deviation– Skewness– Kurtosis– Area– Compactness– Eccentricity– Etc…
Materials & Methods
2005 RDMI Lab Conference
1. Final Extraction Image
2005 RDMI Lab Conference
2. GGO Enhanced Image
2005 RDMI Lab Conference
2. GGO Enhanced Image
2005 RDMI Lab Conference
3. Original CT Image
2005 RDMI Lab Conference
3. Original CT Image
2005 RDMI Lab Conference
Texture AnalysisMaterials & Methods
2005 RDMI Lab Conference
Support Vector Machine
• 11 parameters, 29 cases• Using OSU LIBSVM in matlab• Kernel Type– Polynomial, degree:5
• [AlphaY, SVs, Bias, Parameters, nSV, nLabel]= u_PolySVC(T_Samples, T_Labels, Degree);
• [Labels, DecisionValue]= SVMClass(T_Test, AlphaY, SVs, Bias, Parameters, nSV, nLabel);
• Result– Label : [0 0 1 1]– DecisionValue
Materials & Methods
2005 RDMI Lab Conference
Results
• Image processing
• Texture analysis
completed!
• SVM classification
진행중 ...
• Final Results
2005 RDMI Lab Conference
Further Study
• SVM Training, Testing– 많은 Case, – Training set 과는 다른 Test set 적용
• 통계처리– Sensitivity, Specificity, – ROC curve analysis
• GUI development
Thank you!!!