NA-MIC National Alliance for Medical Image Computing Big Science in Schizophrenia Research Ron Kikinis, M.D. Professor of Radiology, Harvard Medical School, Director, Surgical Planning Laboratory Brigham and Women’s Hospital
National Alliance for Medical Image Computing Co-authors Ron Kikinis, M.D. 1, Tina Kapur, Ph.D. 2, Martha E. Shenton, Ph.D. 1, Jeffrey S. Grethe, Ph.D. 3, Mark H. Ellisman, Ph.D. 3 1 Brigham and Women’s Hospital, Boston, MA, 2 Epiphaniymedical, Seattle, WA, 3 University of California San Diego, La Jolla, CA Acknowledgements: NIH roadmap, NCRR, NIBIB, NCI, NLM, NSF, CIMIT
National Alliance for Medical Image Computing Introduction Schizophrenia research is still on the search for a diagnosis based on quantitative methods Imaging is complementing clinical assessments with subtle findings that are only significant in group comparisons Most schizophrenia research groups have small numbers of subjects and compete fiercely with their peers.
National Alliance for Medical Image Computing Big Science Potential Advantages: –Improved signal through larger subject numbers –Reduced noise through standardization –Potential for new, subtle findings Disadvantages –Clinical and technical challenges –No history of large scale collaboration in the field of schizophrenia research
National Alliance for Medical Image Computing Problems to be Addressed 1.Clinical assessment 2.Patient populations 3.Treatment variations 4.Variations in assessment of genetics 5.Imaging 6.NIH review process
National Alliance for Medical Image Computing 1. Clinical Assessment Diagnostic and clinical symptom measures often vary across sites Solution: Standardization of –patient recruiting, –data acquisition protocols –clinical assessments –fBIRN is working on this problem in the context of schizophrenia
National Alliance for Medical Image Computing 2. Patient Populations Patient populations may vary across sites (e.g. chronic patients versus first episode subjects) Solution: Homogeneous patient groupings based on: – severity of illness, –duration of illness, –chronic versus first episode, etc.
National Alliance for Medical Image Computing 3. Treatment Variation Medication practices may be different at different sites – this also may impact findings and needs to be evaluated. Solution: Algorithms for the administration of medication need to be agreed upon across multiple-centers.
National Alliance for Medical Image Computing 4. Assessment of genetics Genetic measures may be evaluated differently across sites. Solution: Some consensus or plan is needed across sites regarding specific measurements to be used to evaluate genetic contributions to schizophrenia.
National Alliance for Medical Image Computing 5. Imaging Imaging techniques may vary across sites and thus some effort would need to be made to ensure compatibility and comparability across sites. Solution: Standardize a number of imaging protocols that can be used across sites. A great deal of work has already been done here as part of BIRN and related projects.
National Alliance for Medical Image Computing 6. NIH Review process How should the peer review process be modified to maximize excellence in science along with big science? Solution: R01 grant applications and large science enterprises need to be seen as complementary and not competing with one another.
National Alliance for Medical Image Computing Example: Imaging Research Schizophrenia-imaging related research topics –DTI –fMRI
National Alliance for Medical Image Computing Harvard/MIT: DTI Psychiatry Neuroimaging Laboratory
National Alliance for Medical Image Computing Fiber Clustering Left and Right Fornix Uncinate Fasciculus and Inferior Occipito-Frontal Fasciculus Splenium of the Corpus Callosum [O’Donnell L, Kubicki M, Shenton, ME, Dreusicke M, Grimson E, Westin, CF: A Method for Clustering White Matter Fiber Tracts. Am J Neuroradiol (In Press)]. Clustering algorithm: Takes traced fibers (left), extracts features from these fibers (middle), and produces a segmentation based on the similarity (right).
National Alliance for Medical Image Computing Population Comparison Automatic generation of white matter fiber bundles based on shape similarity across subjects. [O’Donnell, et al., MIT] Cingulum BundlesUncinate FasciculiCorpus Callosum
National Alliance for Medical Image Computing Dartmouth: fMRI Andrew J. Saykin, Psy.D., ABPP-CN John D. West, M.S. Robert M. Roth, Ph.D. Brain Imaging Laboratory Dartmouth Medical School / DHMC
National Alliance for Medical Image Computing Functional MRI – Behavioral Performance Verbal Encoding/Retrieval Episodic Memory Task Interim analysis
National Alliance for Medical Image Computing Functional MRI FMRI Activation During Continuous Auditory-Verbal Memory (new > old word contrast) in Patients with Schizophrenia Patients N = 7 p =.01 Interim analysis Verbal Encoding/Retrieval Episodic Memory Task
National Alliance for Medical Image Computing Standardization Clinical measures Image acquisition protocol
National Alliance for Medical Image Computing Technological Infrastructure Data Sharing: BIRN Clinical Measures: fBIRN Data Analysis: NA-MIC
National Alliance for Medical Image Computing Example 1: BIRN CC Building of shared infrastructure –Distributed file and database system (SRB) –Managed Authentication System with single login –Access to compute facilities (teragrid) –Access to shared data
National Alliance for Medical Image Computing BIRN Biomedical Informatics Research Network –Improve Multi-Site Clinical Research –Calibration –Informatics
National Alliance for Medical Image Computing mBIRN Federated Database Cortical Summary Data by Region Subcortical Summary Data by Region
National Alliance for Medical Image Computing BIRN Portal Web Based –Single Login to BIRN Resources –Intuitive Interface –Flexible to Add Tools –Launch Local Visualization Tools on Downloaded Data
National Alliance for Medical Image Computing Example 2: fBIRN “calibration” First large scale attempt at standardization for multi-site fMRI acquisitions. Acquistion protocol on equipment from multiple vendors Calibraiton phantom for technical level calibration Traveling subjects for repeat fMRI studies Large variabilities found Second run in preparation
National Alliance for Medical Image Computing Function BIRN Overview Calibration Methods for Multi-Site fMRI – Study Regional Brain Dysfunction and Correlated Morphological Differences – Progression and Treatment of Schizophrenia Human Phantom Trials – Common Consortium Protocol – 5 Subjects Scanned at All 11 Sites – Additional 15 Controls, 15 Schizophrenics Per Site Per Year Develop Interoperable Post-Processing UC Irvine, UCLA, UC San Diego, MGH, BWH, Stanford, UMinnesota, UIowa, UNew Mexico, Duke/UNorth Carolina, MIT
National Alliance for Medical Image Computing sub106.2 BH2 sub106.4 BH2 Breath-Holding Task
National Alliance for Medical Image Computing Comparing Apples and Oranges Bad News: Different scanners = different raw images Good News The errors are systematic
National Alliance for Medical Image Computing Results: Images Average across 5 individuals at same site, same visit
National Alliance for Medical Image Computing STAPLE Results Site vs. Subject –Subject Variability Greater than Site Variability Field Strength –3T and 4T Detect More Activation than 1.5T –3T and 4T have Less Variability in Sensitivity and Specificity Visits –Less Activation, but More Robust and Systematic Activation in Visit 2 vs. Visit 1
National Alliance for Medical Image Computing Example 3: NA-MIC kit NA-MIC aims at developing image analysis technology for –software developers and –medical researchers Applications: 3D Slicer Software methodologies and software tools: –Multiplatform, nightly builds, automated testing –VTK, ITK
National Alliance for Medical Image Computing NA-MIC: a National Alliance 10 entities develop BSD style open source technology: –MIT, UNC, UTAH, Georgia Tech, MGH –GE GR, Kitware, Isomics, UCSD, UCLA 4 groups act as driving biological projects and use the technology –BWH, Dartmouth, UCI, Toronto Significant outreach activities: –Service, dissemination, training –For example google ‘slicer 101 training’
National Alliance for Medical Image Computing Structure
National Alliance for Medical Image Computing MIT InwardOutward Atlas-based Segmentation Principle Modes of Variation Train Deformable Model
National Alliance for Medical Image Computing Someone broke the build! GE
National Alliance for Medical Image Computing Isomics
National Alliance for Medical Image Computing Dissemination: Events
National Alliance for Medical Image Computing Conclusion Pros –“Big Science” done right is a force multiplier –Allows development and adoption of best practices in research –Faster and higher quality dissemination of new techniques and of new science Cons –Re-education of scientist is necessary and painful –Infrastructure is difficult to explain/justify, because of long lead times between creation and impact