1 / 34

Integrating Machine Learning and Physician Knowledge to Improve the Accuracy of Breast Biopsy

Integrating Machine Learning and Physician Knowledge to Improve the Accuracy of Breast Biopsy. Inês Dutra University of Porto, CRACS & INESC-Porto LA Houssam Nassif , David Page , Jude Shavlik , Roberta Strigel , Yirong Wu , Mai Elezabi and Elizabeth Burnside UW-Madison.

tyrell
Download Presentation

Integrating Machine Learning and Physician Knowledge to Improve the Accuracy of Breast Biopsy

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Integrating Machine Learning and Physician Knowledge to Improve the Accuracy of Breast Biopsy Inês Dutra Universityof Porto, CRACS & INESC-Porto LA HoussamNassif, David Page, JudeShavlik, Roberta Strigel, YirongWu, MaiElezabiandElizabethBurnside UW-Madison

  2. IntroandMotivation • USA: • 1 woman dies of breast cancer every 13 minutes • In 2011: • estimated 230,480 newcases ofinvasivecancer • 39,520 women are expected to die • Source: U.S. Breast Cancer Statistics, 2011 AMIA 2011

  3. IntroandMotivation • Portugal: • Per year: • 4,500 new cases • 1,500 deaths (33%) • Source: Liga Portuguesa Contra o Cancro - September 2011 AMIA 2011

  4. IntroandMotivation • Mammography • MRI • Ultrasound • Biopsies AMIA 2011

  5. IntroandMotivation • Image-guidedpercutaneous core needlebiopsy • standard for thediagnosisofsuspiciousfindings • Over 700,000 womenundergobreast core biopsyinthe US • Imperfect: biopsiescanbeinconclusivein5-15% of cases • ≈ 35,000-105,000 oftheabovewillrequireadditionalbiopsies AMIA 2011

  6. IntroandMotivation • Breastbiopsies: • Fine needlebreastbiopsy • Stereotacticbreastbiopsy • Ultrasound-guided core biopsy • Excisionalbiopsy • MRI-guided AMIA 2011

  7. Objectives • Hypotheses • Canwelearncharacteristicsofinconclusivebiopsies? • Canweimprovetheclassiferbyaddingexpertknowledge? AMIA 2011

  8. Terminology • Inconclusivebiopsies: “non-definitive”: • Discordantbiopsies • Highrisklesions • Atypicalductalhyperplasia • Lobular carcinoma insitu • Radial scar • etc. • Insufficientsampling AMIA 2011

  9. MaterialsandMethods • Data collected • Oct1st, 2005 to Dec 31st, 2008 • Tools • WEKA • Aleph • Validation: leave-one-outcrossvalidation • Resultstested for statisticalsignificance AMIA 2011

  10. Data source 1384 imageguided core biopsies 349 M 1035 B 925 (C) 110 (NC) 43 (D) 51 (ARS) 16 (I) Aftereliminatingredundancies 94 AMIA 2011

  11. Data distribution • 94 non-concordant cases wentthroughadditionalprocedures (e.g., excision) • 15 were “upgraded” frombenign to malignant • Ourpopulation: 15+/79- • Task: find a distinctionbetween upgrades andnon-upgrades AMIA 2011

  12. Data Features AMIA 2011

  13. WEKA andAleph • Data cleaning, singletable • 15-fold stratifiedcross-validation (leave-one-out) • WEKA (BayesTAN) andAleph • Withoutexpertknowledge • Withexpertknowledge AMIA 2011

  14. WEKA andAleph • Waikato Environment for KnowledgeAnalysis • Collectionofmachinelearningalgorithmsand data miningtasks • use “propositionalized” data (flattable) • ALearningEngine for ProposingHypotheses • InductiveLogicProgramming (ILP) system • Knowledgerepresentation: firstorderrules AMIA 2011

  15. Method Learnrulesabouthowandwhynon-definitive cases are “upgraded” Otherrules (simple) providedby a specialist Combine Newrules AMIA 2011

  16. Modelling upgrades withoutexpertknowledge • Alephlearningonthe 94 cases (15+/79-) • Exampleofrulelearned upgrade(A) IF bxneedlegauge(A,9) AND abndisappeared10(A,0) AND mass01(A,1) AND anotherlesionbxatsametime01(A,0). [4+/0-] AMIA 2011

  17. ExamplesofExpertrules (1) Upgrade increases on stereo if calcifications and ADH are present upgrade(A) IF pathdx(A,atypical_ductal_hyperplasia) AND calcifications(A,’a,f’) AND biopsyprocedure(A, stereo). (2) Upgrade increases on US(Ultra-Sound) if high BIRADS category upgrade(A) IF concordance(A,d) AND mammobirads(A,4B/4C/5). AMIA 2011

  18. Modelling upgrades withexpertknowledge • Exampleofrulelearned upgrade(A) IF numOfSpecimens(A,gt6) AND rule1_2(A) AND rule1_3(A). [3+/1-] rule1_2(A) IF pathdxabbr(A,adh) AND calcifications(A,f). [3+/4-] rule1_3(A) IF pathdxabbr(A,adh) AND calcification_distribution(A,g). [3+/6-] AMIA 2011

  19. Results AMIA 2011

  20. Results AMIA 2011

  21. Results:WEKA versus Alephandhumanrules AMIA 2011

  22. Conclusions • Both WEKA andAlephcanimproveresultswhencombining original data withexpertknowledge (withouttuning) • WEKA improvesonfalse positives andcanbeadjusted to achieveRecallof 100% withtheneed to re-examinearound60 women (outof 79) • Alephimprovesonfalse negatives andproduces a betterclassifierwiththeadvantageofusing a languagethatiseasilyunderstandablebythespecialist AMIA 2011

  23. FutureWork • NLM projectstartedthismonth AMIA 2011

  24. FutureWork • NLM projectstartedthismonth • Shorttermtasks: • Augmentthecurrentdataset • Reviewconsistentlymisclassifiedexamples • Possibily use associationrules to produce a firstsubsetofrules AMIA 2011

  25. FutureWork • NLM projectstartedthismonth • Shorttermtasks: • Augmentthecurrentdataset • Reviewconsistentlymisclassifiedexamples • Possibily use associationrules to produce a firstsubsetofrules • Mediumtermtasks: • learnfromthe cases thatwerenotupgraded • Developtheiterativecycleoflearningandexpertrulerevision AMIA 2011

  26. FutureWork • NLM projectstartedthismonth • Shorttermtasks: • Augmentthecurrentdataset • Reviewconsistentlymisclassifiedexamples • Possibily use associationrules to produce a firstsubsetofrules • Mediumtermtasks: • learnfromthe cases thatwerenotupgraded • Developtheiterativecycleoflearningandexpertrulerevision • Longtermtasks: • Use more sophisticatedlearningmethods to producebetterrules • To integratebiopsyattributesandtheclassifiersobtained to MammoClass (http://cracs.fc.up.pt/pt/mammoclass/) AMIA 2011

  27. Acknowledgments • UW-MadisonMedicalSchool • FCT-PortugalandHorusandDigiScopeprojects • AMIA reviewersandorganizers • Specialack to theUW-MadisonMedicalSchoolpeoplethatwerenever “scared” ofcomputertechnologyandhavebeenalwaysveryenthusiasticaboutusingcomputationalmethodologies to helptheirwork AMIA 2011

  28. CONTACT: ines@dcc.fc.up.pt THANK YOU! AMIA 2011

  29. Experiment 1Significancetests - p-values AMIA 2011

  30. Experiment 2tuning • 5x4 stratifiedcrossvalidation • Parameterstuned: • Minacc: 0.05 and 0.1 • Minpos: 1, 2, 3 • Noise: 0, 1, 2 AMIA 2011

  31. Experiment 2tuning, bestparametersperfold AMIA 2011

  32. Experiment 2 - Results p= 0.1 0.3 0.03 AMIA 2011

  33. Summary AMIA 2011

  34. UW upgrades data • 94 benign cases afterbiopsieswerediscussed • 15 considered to be, infact, malignant upgrades • Tasks: • learncharacteristicsof upgrades that do notappearinnon-upgrades interpretablerules • Canweimprovetheclassiferbyaddingexpertknowledge? AMIA 2011

More Related