2013 machine learning_choih
- 1. ARTIFICIAL INTELIGENCE in Nuclear Medicine The latest trend of statistical approach for ‘Big Data’ in medicine R3, Choi Hongyoon
- 2. CONTENTS INTRODUCTION Why A.I. for medicine? New Statistical Approach REVIEWS Multivariate analysis for ‘prediction’ ACTUAL PRACTICE Real Application New Design for PET/MR studies ARTIFICIAL INTELIGENCE in Nuclear Medicine
- 6. INTRODUCTION
- 9. INTRODUCTION Image Variables Pattern Database Results SUV, Clinical Factors, Texture, etc. Diagnosis (CAD) Survival / Morbidity / Prediction of Tx response Previous Data : EMR, PACS , …
- 10. INTRODUCTION TREND of medical statistics Chance 2012;25(3):31-34
- 11. INTRODUCTION Benign vs. Malignancy SUV : p = 0.23 Size : p = 0.43 TLG : p = 0.17 MTV : p = 0.22 CEA : p = 0.44
- 12. Prediction From Big DATA Statistical methods and Clinical ideas
- 13. REVIEW Application of machine learning in medicine Lung nodule characteristics/ Clinical factors Logistic Regression (Training 7008 nodules/ Validation 5021 nodules) Malignancy Prediction : ROC 0.90 McWilliams et al. NEJM, 2013.
- 14. REVIEW Application of machine learning in medicine Malignancy Prediction : ROC 0.90 McWilliams et al. NEJM, 2013. Logistic Lung nodules Size Age, Sex Location Character Predicting Lung Cancer
- 15. REVIEW Application of machine learning in medicine • Images • Genetic Data (SNPs, Sequencing,…) • Clinical Features Ingredient • Regression (e.g. Logistic) • Kernel (e.g.)Support Vector Machine • Artificial Neural Network • Ensamble (e.g. Random Forest) Methods • Diagnosis • Survival • Treatment Prediction • Imaging : Segmentation / Registration Results
- 16. REVIEW Arsanjani R. et al. Improved Accuracy of Myocardial Perfusion SPECT for the Detection of Coronary Artery Disease Using a Support Vector Machine Algorithm. J Nucl Med 2013. Prels O. et al. Neural Network Evaluation of PET Scans of the Liver: A Potentially Useful Adjunct in Clinical Interpretation. Radiology 2011.
- 17. SVM REVIEW Myocardial SPECT for Dx. of CAD Automatic Softwares No diagnostic score based on multiple quantitative features Support Vector Machine (SVM) Kernel-based machine learning algorithm SVM for Myocardial SPECT Functional MPS Variables (EF, SDS, …) Predicting Severe Stenosis
- 18. REVIEW Support Vector Machine Non-linear classification (High-dimensional Kernel based) Hyperplane (Decision-plane) SVM for Myocardial SPECT EF Thikcening High- Risk Low- Risk Training Test Patient is High-Risk Group Simulated on Libsvm, http://www.csie.ntu.edu.tw/~cjlin/libsvm/
- 19. REVIEW Support Vector Machine SVM for Myocardial SPECT Nonlinear Hyperplane Kernel Based (Higher Dimension) Logistic Regrssion Error ~40% SVM Error ~7% Simulated on Libsvm, http://www.csie.ntu.edu.tw/~cjlin/libsvm/
- 20. REVIEW METHODS 957 Myocardial SPECT studies Training Group (n=125) : 25 LLk, 25 0-vessel, 25 1-vessel, 25 2-vessel, 25 3-vessel Testing Group (n=832) Myocardial SPECT : Tc-99m sestamibi rest/stress Stress Total Perfusion Deficit (TPD) Ischemic Changes (ISCHs) Poststress EF changes (EFCs) Motion and thickening changes (MTC) SVM for Myocardial SPECT
- 21. REVIEW METHODS Visual Scoring : 2 Board NM physicians / SDS scoring Support Vector Machine Training : TPD, ISCHs, EFC or MTC combination Polynomial kernel Testing : Probability / CAD vs. non-CAD prediction Gold Standard : LMA >50% or Other Coronary a. >70% stenosis SVM for Myocardial SPECT
- 22. REVIEW RESULTS SVM for Myocardial SPECT Combination Methods >> Single Feature
- 23. REVIEW SVM for Myocardial SPECT RESULTS Combination Machine Learning > or = Boardman’s Reading
- 24. REVIEW SVM for Myocardial SPECT CONCLUSION Improvement of diagnostic performance using multiple functional & perfusion parameters Full automated risk evaluation
- 25. REVIEW Arsanjani R. et al. Improved Accuracy of Myocardial Perfusion SPECT for the Detection of Coronary Artery Disease Using a Support Vector Machine Algorithm. J Nucl Med 2013. Prels O. et al. Neural Network Evaluation of PET Scans of the Liver: A Potentially Useful Adjunct in Clinical Interpretation. Radiology 2011.
- 26. REVIEW FDG PET for hepatic metastasis Sensitivity : 86% MR vs. 71% FDG-PET Variability of hepatic metastases & heterogeneous liver parenchymal uptake ANN for FDG PET of the Liver ANN FDG PET Variables (SUVs, Clinical Factors) Predicting Liver Metastasis /Compared with MR
- 27. REVIEW Artificial neural network ANN for FDG PET of the Liver Neuron Perceptron SUV > 2.5 Size > 1cm g(x) Liver SUV < 3.0 (x30) (x10) (x-10) >35 : 1 <35 : 0 Benign Node Hidden Layer Output
- 28. REVIEW Artificial neural network ANN for FDG PET of the Liver SUV > 2.5 Size > 1cm Liver SUV < 3.0 A … g(x) g(x) g(x) g(x) g(x) Metast asis Training Set Weighting Factors Maximize correct output Testing Set Range 0 - 1
- 29. REVIEW METHODS Patients: 98 FDG PET scans and liver MR Input Nodes Lesion indepdent : Ages, Liver SUV&SD, spleen SUV SD, gluteus maximus SUV SD, BMI, glucose level (9 nodes) Lesion specific: lesion SUV & SD (2 nodes) Neural Network 11 nodes Hidden layer 5 (arbitrarily) 1 Output ANN for FDG PET of the Liver
- 30. REVIEW METHODS Physician Performance Two NM/abdominal imaging experties Blind to NN data / Unblind to NN data Reference Standard Biopsy + followup MR ANN for FDG PET of the Liver
- 31. REVIEW RESULTS ANN for FDG PET of the Liver Lesion Depedent newtork (AUC 0.905) Observer 1 (0.786 0.924) Blind unblind Observer 2 (0.796 0.881) Blind unblind
- 32. REVIEW ANN for FDG PET of the Liver RESULTS May 2008 ANN score 0.70 August 2008 ANN score 0.88 December 2008 ANN score 1.00
- 33. REVIEW CONCLUSION Lesion independent nodes 10 cases : absent of visually apparent lesions mets Interpretation of systemic / global abnormalities Reinterpretation after NN Significantly improved visual reading ANN for FDG PET of the Liver
- 34. REVIEW Applications Thyroid nodule : Indeterminate nodule Applications NEJM, 2012
- 35. REVIEW Applications 49 Clinical sites, 4812 FNAB 577 Indeterminate nodules Gene-expression Classifier (167 genes, SVM-based) Chudova D, et al. J Clin Endo Metab 2010. Applications NEJM, 2012 92% sensitivity 52% specificity
- 36. REVIEW Applications Applications J Clin Oncol 2009 SVM Several Biomarkers for Lung Cancer Survival
- 37. REVIEW Applications 148 Stage IB NSCLC, IHC results 73 training set + 75 validation set 5Yr survival 1 ; death 0 Different Classifier : SVM1, 2, 3 different IHC features Applications J Clin Oncol 2009
- 38. REVIEW Applications Applications J Clin Oncol 2009 Validation Set
- 39. REVIEW Benign vs. Malignancy SUV : p = 0.23 Size : p = 0.43 TLG : p = 0.17 MTV : p = 0.22 CEA : p = 0.44 Applications
- 40. REVIEW Applications Brain Studies Applications Zhang D, et al. NeuroImage 2012. SVM PET MRI CSF AD vs. Normal Template – based multiple VOIs
- 41. REVIEW Applications Brain Studies Applications Gray KR, et al. NeuroImage 2013. Random -Forest PET MRI CSF AD vs. Normal Random Forest
- 42. REVIEW Applications Brain Studies Applications Important Areas Gray KR, et al. NeuroImage 2013. Combined Random Forest : Se 87.9% ; Sp 90.0%
- 43. Actual data – Simple Experience Difficult, but we can do it.
- 44. ACTUAL DATA Simple Experience Pre-CCRT locally advanced rectal cancer From NCC CCRT Or
- 45. ACTUAL DATA J Nucl Med 2011. J Nucl Med 2012. Lung cancer Esop cancer
- 46. ACTUAL DATA Simple Experience Pre-CCRT locally advanced rectal cancer From NCC Tumor Segmentation (SUV threshold, Isocontour) Image-based Features SUVmax, SUVmean, Kurtosis, Skewness Textural Features: Contrast, Correlation, Entropy, Homogeneity, Energy Prediction for Dworak Tumor regression grade
- 47. ACTUAL DATA Simple Experience Pre-CCRT locally advanced rectal cancer From NCC Hypothesis : Different Textural Features in Good responders T-tests for each of features b/w good and poor responders SUVmax 8.2 SUVmean 7.5 Kurtosis 1.3 …
- 49. ACTUAL DATA Simple Experience 52 Training set + 20 Test set Image Features – SVM Prediction for CCRT response Features Subjects Response
- 50. ACTUAL DATA Simple Experience >> svmStruct=svmtrain(trainingData,response,’kernel_function’,’rbf’); >> PredictResponse=svmclassify(svmStruct,testData); Test DATA (n=20) Results : 17/20 Accurate Prediction for Response
- 51. Summary MULTIPLE Data Random or nonrandom (pattern?) ‘Arrival Time of E-mails’ Non-random variables Rule of specific output Benign vs. Malignancy SUV : p = 0.23 Size : p = 0.43 TLG : p = 0.17 MTV : p = 0.22 CEA : p = 0.44 SUV, Size, TLG , … : Unknown Rule : Find Rule!
- 52. SUMMARY Image Variables Pattern Database Results SUV, Clinical Factors, Texture, etc. Diagnosis (CAD) Survival / Morbidity / Prediction of Tx response Previous Data : EMR, PACS , …
- 53. SUMMARY PET/MR studies PET variables and MR variables Not competitive e.g.> SUVmax, SUVmean, TLG, MTV ADC, perfusion MR parameters, Size, … Combination Studies Big Data analysis
- 54. Take Home Message Multiple variables based prediction. State-of-the-art statistical approach PET/MR Image ‘big data’ Integration using new methods for better diagnostic performance A wide range for application / Not so difficult ANOVA, T-tests, nonparametric tests, …
Editor's Notes
- #12: 조금 더 가까이 와닿게 말씀드릴 수 있는 주제라고 한다면…..
- #16: 우선 두 개의 우리 과와 관련된 논문을 Review하고 쓰임새나 활용범위에 대해 다른 논문들도 간단하게 소개하도록 하겠다..
- #17: Cedars Sinai에서 나온 논문..
- #20: EF, Thickness라면 이 둘의 vector 곱을 통해 2차원을 3차원으로 만들어준다..
- #26: From MGH
- #40: 조금 더 가까이 와닿게 말씀드릴 수 있는 주제라고 한다면…..
- #41: Out of 관심… 이겠으나, Brain 을 다음과 같이 분석하는 방법도 행해지고 있다. Network 분석에도 적용할 수 있고, 기존의 Raw data로도 적용할 수 있는 Tool로서 neuroImage등에서 다변화되어 나오고 있음.
- #42: Out of 관심… 이겠으나, Brain 을 다음과 같이 분석하는 방법도 행해지고 있다. Network 분석에도 적용할 수 있고, 기존의 Raw data로도 적용할 수 있는 Tool로서 neuroImage등에서 다변화되어 나오고 있음.
- #43: ‘문항에 들어갈 수 있는 영역들’
- #51: Kernel 조정을 어떻게 하는가와 향후 Cross-validation이라 하여 Training set 내에서의 변수를 어떻게 조정하는가 등의 절차로 최적화하는 방법이 남아있지만, 대략적으로 다음과 같은 흐름으로 결과를 얻을 수는 있다는 것 !!!
- #52: 과연 Email의 도착시간은 랜덤할까? ; 나에게 어떤 행동반경이 있을 것이며 multiple 한 factor의 영향을 받는다. 역으로 그를 추적한다면, 내가 어디사는지 (시차로부터), 내 직업이 무엇인지 (Database로부터..), …. 저 Data들은 결코 랜덤하지 않다. 어떤 규칙이있을 것이며 그를 역으로 추적한다면 원인을 파악할 수 있다. 우리가 가진 Data가 하나가 아닌 여러 개일 때 그를 만들어낸 Core를 찾아내기 위한 노력을 한다면 얼마든지 가능하며 앞의 통계적인, 컴퓨터적인 기법을 통해 충분히 가능할 것이다.
- #53: 지극히 Computation 된 방법으로 이렇게 진단에 도움을 받는 것이 꼭 기술을 극단적으로 이용하는 것은 아닐 것이다. 오히려 이러한 것들이 신뢰할 수 있을 정도로 환자의 Mortality를 반영하고 Morbidity를 반영하여 어떤 조언을 해줄 수 있다고 한다면, 이는 Computer를 활용하였지만 어떤 것보다도 더 Humanized된 기술이 될 수 있을 것이다. 환자에게 SUV를 설명하고 borderline malignancy하다는 것을 알려주는 것이 무슨 의미가 있는가를 다시 곱씹어 볼때가 아닌가 한다.
- #55: ANOVA, T-Tests 등등 원리를 완벽히 이해하고 통계를 돌리는 MD 가 몇 명이나 될 것인가..