Machine Learning Powered Medical Image Analytics: Tools Waiting for Applications George Washko MD Division of Pulmonary and Critical Care Medicine Brigham and Women’s Hospital
Disclosures • Consultancies: Astra Zeneca, GlaxoSmithKline, Boehringer Ingelheim, BTG, Genentech, Janssen, Philips, PulmonX, Novartis, Regeneron, Vertex • Quantitative Imaging Solutions: Founder of a consulting group and software development LLC for data management • Grants: NIH, Boehringer Ingelheim, BTG (EKOS), Janssen, Lung Biotechnology • Spouse: Works for Biogen
Medical Applications of ML Matching Engines Diagnosis and Treatment Patients with similar profile Clinical Decision Support Treatments with similar Symptoms analyzer Cost-Benefit ratio Automated Radiology Treatment efficacy Discovery Virtual Assistant Clinical Trials Smart telemedicine Hypothesis generation Doctor assistants ML Proof of concept Healthcare Precision Medicine Workflows Genomics Patient flow optimization Tailored therapies Detect process inefficiencies Predictive Modeling Mobile Apps Risk stratification Well-being Healthcare Analytics Chronic disease management
ACIL Activities Framingham Heart COPDGene Study (PRC) •10,000 subjects with HRCT •Genetic epidemiology of • Longitudinal analysis of lung disease COPD •Population-based (3000 subjects) CT Analysis Applied Chest Imaging Laboratory NETT CARDIA • 1,200 with severe COPD • Coronary Artery Risk Development • Extensive phenotypic data in Young Adults •Population based study Clinical Data
Why Thoracic Imaging? • Asthma • Airways • BOS • Vasculature • Bronchiectasis • Parenchyma • CVD • Heart • COPD • Muscle • ILD • Bone • Metabolic • Adipose tissue syndrome • Bone marrow • Lung Cancer • PVD/VTE
What’s Possible with ML Powered Image Analytics? Image navigation/curation Disease and comorbid condition • Detection • Stratification • Prognostication – Natural history – Therapeutic response (“patients like me”) • Monitoring therapeutic response
Organ Detection Onieva, SPIE, 2018
Quantification without Segmentation Bone Mineral Density González G, SPIE 2018
Organ Segmentation
RV (blue) and LV (red) Elevated PA pressures Systolic dysfunction and Normal with RV dilation dilation of the LV Acad Radiol. 2017;24(5):594-602.
Predicting Clinical Outcomes Acad Radiol. 2017;24(5):594-602.
1157 HRCT of fibrotic lung disease Comparison of algorithm performance to Training: 929, Validation: 89, Test: 139 majority vote of 91 specialists Walsh et al. Lancet Resp Med 2018 .
Direct Outcome Estimation • Respiratory Events • Hospitalization Input: 512x512x1 C1: 128x128x16 C2: 32x32x32 C3: 8x8x64 FC1: 1024 FC2: {2,#classes,1} + MaxPooling + MaxPooling +MaxPooling 64 Conv Conv Conv 32 16 González, Am J Respir Crit Care Med, 2017
Path Forward? Opportunities for ML powered image analytics: – Clinical care • Bringing expertise to resource limited areas (ILD, Lung Cancer, etc) – QI – Billing/resource allocation • Challenges: – Model Corruption and Regulatory Review – Who will champion the effort?
COPDGene • 10K smokers with a range of lung dysfunction • Volumetric CT scan • 60% NHW, 40% AA
Magic? • OR 38.9! • Ethnicity – How did it work? – Well documented differences in body composition/fat free mass – Skeletal structure
COPDGene Data Revisited NHW AA
Bias and Discrimination • Black and Latinx patients had 9% and 17% lower rates of admission to the cardiology service
Acknowledgements • Raul San Jose Estepar • George Washko • David Bermejo-Pelaez • Samuel Ash • Ruben San Jose Estepar • Angela Blake • Sarah Gerard • Carolyn Come • Monica Iturrioz • Alejandro Diaz • Rafael Moreta • Wocizch (Remy) Dolliver • Pietro Nardelli • Stefanie Mason • James Ross • Carrie Pistenmaa • Gonzalo Vegas Sanchez-Ferrero • Nick Rahaghi • German Gonzalez Serrano • Andrew Tsao • NHLBI • DoD • Boehringer Ingelheim • Johnson and Johnson • Lung Biotechnology • BTG Therapeutics
Recommend
More recommend
