Segasist TM Overcoming Variability in Medical Image Contouring Dr. Farzad Khalvati – Chief Technology Officer farzad.khalvati@segasist.com March 2012 www.segasist.com ebo-enterprises.com Contouring Region of interest (ROI), e.g. tumour Contouring by clinician (radiologist, oncologist, pathologist etc.) Any medical image (CT/MRI/US/PET etc.) Copyright © Segasist Technologies 2008-11 Contouring is necessary • Cancer treatment needs contouring • Cancer occurs frequently; e.g. Prostate cancer: • The most common non-skin cancer for adult males • The third leading cause of cancer death for men in Canada with incidence rates on the rise • One in six men in Canada will be afflicted by prostate cancer during their lifetimes. • Contouring is an important part of diagnosis, monitoring, and treatment -3Copyright © Segasist Technologies 2008-11 Contouring: The Challenge of Segmentation Many modalities/cases: 664 Billion images/year in the US alone Prostate MR Breast U/S Brain CT Prostate U/S Lung X-Ray Software A Software B Software C Software D Software E Extracted lesion/tissue/organ used for diagnosis/treatment planning/intervention -4Copyright © Segasist Technologies 2008-11 Small Problem: -5Copyright © Segasist Technologies 2008-11 Small Problem: Contouring takes time -6Copyright © Segasist Technologies 2008-11 Demand Snapshot: Radiation Oncology Volume Contouring Dose Calculation Treatment Copyright © Segasist Technologies 2009-11 Contouring is a major bottleneck (0.25-3 hours/patient) 2010-2020: The number of cancer patients will increase by 22%, while the number of radiation oncologists will increase by just 2%. Study published in The Journal of Clinical Oncology, October 18, 2010 7 Bigger Problem: Experts contour differently Contouring is qualitative…. First expert Second expert Inter-Observer Variability First expert Second expert -8Copyright © Segasist Technologies 2008-11 Biggest Problem: Same expert contours differently Contouring is qualitative…. First expert First expert contours again Intra-Observer Variability First expert First expert contours again -9Copyright © Segasist Technologies 2008-11 Inter- and Intra-Observer Variability "The failure by the observer to measure or identify a phenomenon accurately, which results in an error. Sources for this may be due to the observer's missing an abnormality, or to faulty technique resulting in incorrect test measurement, or to misinterpretation of the data." Source: National Library of Medicine Inherent anatomical vagueness/ambiguity Limitations of imaging devices Level of expertise of the expert (Partial) Subjectivity - 10 Copyright © Segasist Technologies 2008-11 The Curse of Variability: Solution • There is no Perfect segmentation algorithm • Consensus Contour: for a given organ/tumour, consensus contour of multiple contours is the one that agrees with all of them the most • Different algorithms can be used: STAPLE • The result contour has maximum sensitivity and specificity with all input contours - 11 Copyright © Segasist Technologies 2008-11 The Curse of Variability: Examples Soft-tissue sarcoma: 13% [Roberge et al., Cancer/Radiothérapie 2011] Prostate: 18% [White et al., Clinical Oncology 2009] Bladder: 32% [Foroudi et al., Med. Imaging & Rad. Onc., 2009] Abdominal aorta: 40% [England et al., Radiography 2008] Breast lumpectomy cavity: 45% [Dzhugashvili et al., Rad.Onc. 2009] Pulmonary nodules: 54% [Bogot et al., Academic Radiology 2005] … - 12 Copyright © Segasist Technologies 2008-11 Conventional Consensus Building It requires experts actually contour the same image Not feasible: Too costly to afford! - 13 Copyright © Segasist Technologies 2008-11 Semi-Conventional Consensus Building • Instead of experts actually contour the same image; • Use previously created Atlases of the experts to generate contours • Use the Atlas-based generated contours to build consensus - 14 Copyright © Segasist Technologies 2008-11 Conventional Atlas-Based Segmentation Atlas New Image Best Match - 15 Copyright © Segasist Technologies 2008-11 Registration Consensus Building Average Weighted average Distance optimization STAPLE algorithm - 16 Copyright © Segasist Technologies 2008-11 Segasist Reconcillio Variability captured One user All users Consistency verification Consensus building intra-observer variability inter-observer variability - 17 - Copyright © Segasist Technologies 2008-11 Segasist Reconcillio Computational Consensus - 18 Copyright © Segasist Technologies 2008-11 Segasist Technologies • University of Waterloo Spin-off • Founded in 2008 • Toronto-based • Products: • Prostate Auto-Contouring: FDA cleared • Segasist Auto-Contouring • Segasist Reconcillio - 19 Copyright © Segasist Technologies 2008-11 Segasist TM Thank You Questions? Dr. Farzad Khalvati, Ph.D. – Chief Technology Officer farzad.khalvati@segasist.com www.segasist.com