View Learning: An extension to SRL An application in MammographyBackgroundThe ProblemCommon Mammography findingsCalcificationsMassArchitectural distortionOther important featuresOther variables influence riskStandardization of PracticeSlide 11Mammography DatabaseOriginal Expert StructureSlide 14Types of LearningLevel 1: ParametersLevel 2: StructureSlide 18Slide 19Slide 20Level 3: AggregatesSlide 22Slide 23Slide 24Level 4: View LearningStructure Learning AlgorithmsNaïve Bayes NetExample TAN NetTANTAN AlgorithmGeneral Bayes NetSparse CandidateSparse Candidate AlgorithmSlide 34Our Initial Approach for Level 4Using ViewsSample RuleMethodologySlide 39EvaluationROC: Level 2 (TAN) vs. Level 1Precision-Recall CurvesSlide 43Slide 44Related Work: ILP for Feature ConstructionWays to Improve PerformanceRicher View Learning ApproachesSlide 48Integrated View/Structure LearningSlide 50Slide 51Slide 52Richer View Learning (Cont.)ConclusionsWILD GroupView Learning: An extension to View Learning: An extension to SRL An application in SRL An application in Mammography Mammography Jesse Davis, Beth Burnside, Inês Dutra Jesse Davis, Beth Burnside, Inês Dutra Vítor Santos Costa, David Page, Jude Vítor Santos Costa, David Page, Jude Shavlik & Raghu RamakrishnanShavlik & Raghu RamakrishnanBackgroundBackgroundBreast cancer is the most common cancerBreast cancer is the most common cancerMammography is the only proven screening testMammography is the only proven screening testAt this time approximately 61% of women have At this time approximately 61% of women have had a mammogram in the last 2 yearshad a mammogram in the last 2 yearsTranslates into 20 million mammograms per Translates into 20 million mammograms per yearyearThe ProblemThe ProblemRadiologists interpret Radiologists interpret mammogramsmammogramsVariability in among radiologists Variability in among radiologists differences in training and differences in training and experienceexperienceExperts have higher cancer Experts have higher cancer detection and less benign biopsiesdetection and less benign biopsiesShortage of expertsShortage of expertsCommon Mammography Common Mammography findingsfindingsMicrocalcificationsMicrocalcificationsMassesMassesArchitectural distortionArchitectural distortionCalcificationsCalcificationsMassMassArchitectural distortionArchitectural distortionOther important featuresOther important featuresMicrocalcificationsMicrocalcificationsShape, distribution, stabilityShape, distribution, stabilityMassesMassesShape, margin, density, size, stabilityShape, margin, density, size, stabilityAssociated findingsAssociated findingsBreast DensityBreast DensityOther variables influence Other variables influence riskrisk•Demographic risk factorsDemographic risk factorsFamily HistoryFamily HistoryHormone therapyHormone therapyAgeAgeStandardization of PracticeStandardization of Practice-Passage of the Mammography Quality Standards -Passage of the Mammography Quality Standards Act (MQSA) in 1992Act (MQSA) in 1992-Requires tracking of patient outcomes through -Requires tracking of patient outcomes through regular audits of mammography interpretations regular audits of mammography interpretations and cases of breast cancerand cases of breast cancer-Standardized lexicon: BI-RADS was developed -Standardized lexicon: BI-RADS was developed incorporating 5 categories that include 43 unique incorporating 5 categories that include 43 unique descriptorsdescriptorsMassDensity-high-equal-low-fat containingShape-round-oval-lobular-irregularMargins-circumscribed-microlobulated-obscured-indistinct-SpiculatedAssociatedFindingsSpecialCasesArchitecturalDistortionCalcificationsHigher ProbabilityMalignancy-pleomorphic-fine/linear/branchingIntermediate-amorphousTypically Benign-skin-vascular-coarse/popcorn-rod-like-round-lucent-centered-eggshell/rim-milk of calcium-suture-dystrophic-punctateBI-RADSTrabecularThickeningSkinThickeningNippleRetractionSkinRetractionSkinLesionAxillaryAdenopathyFocal AssymetricDensityAssymetricBreast TissueLymphNodeTubularDensityDistribution-clustered-linear-segmental-regional-diffuse/scatteredMammography DatabaseMammography DatabaseRadiologist interpretation of Radiologist interpretation of mammogrammammogramPatient may have multiple mammogramsPatient may have multiple mammogramsA mammogram may have multiple A mammogram may have multiple abnormalitiesabnormalitiesExpert defined Bayes net for Expert defined Bayes net for determining whether an abnormality determining whether an abnormality is malignantis malignantOriginal Expert StructureOriginal Expert StructureP1 1 5/02 Spic 0.03 RU4 B P1 2 5/04 Var 0.04 RU4 M P1 3 5/04 Spic 0.04 LL3 B … … … … … … … Patient Abnormality Date Mass Shape … Mass Size Loc Be/MalMammography DatabaseMammography DatabaseTypes of LearningTypes of LearningHierarchy of ‘types’ of learning that Hierarchy of ‘types’ of learning that we can perform on the we can perform on the Mammography databaseMammography databaseLevel 1: ParametersLevel 1: ParametersBe/MalShape SizeGiven: Features (node labels, or fields in database), Data, Bayes net structureLearn: Probabilities. Note: probabilities needed are Pr(Be/Mal), Pr(Shape|Be/Mal), Pr (Size|Be/Mal)Level 2: StructureLevel 2: StructureBe/MalShape SizeGiven: Features, Data Learn: Bayes net structure and probabilities. Note: with this structure, now will need Pr(Size|Shape,Be/Mal) instead of Pr(Size|Be/Mal).P1 1 5/02 Spic 0.03 RU4 B P1 2 5/04 Var 0.04 RU4 M P1 3 5/04 Spic 0.04 LL3 B … … … … … … … Patient Abnormality Date Mass Shape … Mass Size Loc Be/MalMammography DatabaseMammography DatabaseP1 1 5/02 Spic 0.03 RU4 B P1 2 5/04 Var 0.04 RU4 M P1 3 5/04 Spic 0.04 LL3 B … … … … … … … Patient Abnormality Date Mass Shape … Mass Size Loc Be/MalMammography DatabaseMammography DatabaseP1 1 5/02 Spic 0.03