Brain Science Meets Big Science Grid Computing for Clinical Neuroscience Research Steve Pieper, PhD Isomics, Inc. In collaboration with: Surgical Planning Laboratory An NCRR National Resource Center Brigham and Women's Hospital a teaching affiliate of Harvard Medical School http://spl.harvard.edu Acknowledgements Sponsors NIH: NCRR/NCI/NLM, NSF, DOD... SPL F Jolesz, MD, R Kikinis, MD, S Warfield, PhD, W Wells, PhD, C-F Westin PhD, M Halle, PhD, HJ Park, PhD, M Kaus, A Brun, D Kacher Neurosurgery P Black, MD, F Talos, MD MIT E Grimson, PhD, M Ferrant, D Gering, SM, T Kapur, PhD, S Larsen, M Levington, PhD SDF D. Small, M. McKenna http://spl.harvard.edu Outline Why Neuroscience Needs Grid Computing What we do and how we do it How the SPL is Developing Grid Computing BIRN, Neurosurgery, Query Atlas Three Predictions The Evolution of Grid Computing for Neuroscience http://spl.harvard.edu The Challenge Get the Most Societal Benefit from the Investment of Research Time and Resources Need to attract and hold the interest of our clinical collaborators High Quality Tools Production-Oriented Level of Service Need to remove obstacles that prevent adoption – technical and cultural Perception of Local Control Need to Deliver Results http://spl.harvard.edu The SPL – Example Grid Users Algorithm Development Segmentation, Registration, Visualization, Tensor Imaging Software Development 3D Slicer: Free, VTK/ITK-based Medical Applications – neuro and other Image Guided Therapy: Surgical Planning, Intra-operative Assist, Robotics, Follow-up Basic Science Research: Developmental Neuroinformatics, Disease Tracking (MS, AD, Schiz), Digital Atlases http://spl.harvard.edu SPL Collaborations BWH: Radiology, Surgery... Longwood Medical Area: Children's Hospital, Harvard Medical School, Dana Farber Cancer Inst. Boston: MGH (Research and Clinical), MIT, BU, VA, CIMIT National: BIRN, CISST (NSF ERC), GE, Many Individual Collaborations International: Japan (AIST), Tokyo U., Osaka U. Europe (examples) France: INRIA Sophia Antipolis Germany: Hamburg, Heidleburg Sweden: Linkoping Canary Islands Australia: Melbourne, Sidney http://spl.harvard.edu Categories of Research Analysis of Individuals Focus Resources (Computational and Human) on Developing and Performing Treatment In Support of Clinicians Analysis of Populations Processing Pipelines to Extract Derived Data In Support of Neuroscientists Development of Explanations Cross-disciplinary, Multi-scale Quest for Fundamental Understanding http://spl.harvard.edu Basic Neuroanatomy Gray Matter (Cortex) Neuron Cell Bodies Specialized Cortical Regions White Matter Fiber Tracts Ventricles Filled with Cerebro-Spinal Fluid (CSF) Highly Vascularized Brain Responsible for 20% of Metabolism http://spl.harvard.edu Diffusion Tensor Tractography Multiple MR Gradient Acquisitions Three crossing fiber tracts Sensitive to Brownian Diffusion of Water Cell Membranes Restrict Diffusion Post Processing to Extract Probable White Matter Tracts Actual Tracts are Below the Resolution of the Scan http://spl.harvard.edu Provided by Westin, Park, et al Segmentation and Tractography http://spl.harvard.edu Segmentation: MGH; Tractography, Visualization: BWH Developmental Neuroinformatics a).Premature (31 weeks) b).Fullterm equivalent http://spl.harvard.edu c).9 months Unusual to have the opportunity to acquire longitudinal data. Images: Simon Warfield Baby Frontal Lobe Anisotropy 34 Weeks http://spl.harvard.edu 42 Weeks Images: Simon Warfield Integrated Knowledge as Atlas Atlas construction: interactive edit of tissue classification displayed with 3DSlicer. http://spl.harvard.edu Images: Simon Warfield Pre-Operative Map Structural Magnetic Resonance Imaging (MRI) DTI Diffusion Tensor Imaging fMRI Functional MRI MEG Magneto Encephlogram Anatomy Atlas “Textbook” Information http://spl.harvard.edu Grid Projects at the SPL Note: Neuroscience is not Physics or Astronomy Where a physics paper may have hundreds of authors, you need to argue to justify more than 5 or 6 authors on a clinical neuroscience paper Clinical collaborators come from a 'descriptive science' background A new grid paradigm is needed to support the next generation of neuroscience research http://spl.harvard.edu BIRN Calibrated Acquisition Consistent and Well-Curated Archive Raw Data, Human Subject Data, Derived Data Federated Data Grid, Subject Protection Software Interoperability Among Flagship Analysis and Visualization Tools Migration of Key Computation Tasks to Teragrid http://spl.harvard.edu BIRN Challenges Technology Large Data Sets: Storage, Archive, Retrieval Network Infrastructure, Data Grid/Compute Grid, Software Infrastructure Science Multi-site Images: Calibration, Standardization Clinical Scales / Terminology Governance Joint approach to IRB/HIPAA Data Sharing (Ownership, Publications) Sociology: Encouraging Collaborations, IP, Accountability, Credit ... http://spl.harvard.edu BIRN Achievement A Working National Collaboration in a Field without a “Big Science” Culture Even with the best intentions and investigators who really believe in the benefits, sharing is hard http://spl.harvard.edu Morphometry BIRN SRB GE Siemens Picker DICOM Files Philips Sort Images Deidentify Go/NoGo - Standard -Deface or Mask - Clean DICOM Header - Render Movies -Display Movies - QA Approval of Defacing BWH/MGH Duke Directory Hierarchy - Identify Deface Series UCI Conversion to DICOM Retrospective Data Archives (various formats) http://spl.harvard.edu Institutional Firewall Upload -Extract Metadata - Optimize for SRB UCSD Facial Deidentification http://spl.harvard.edu Courtesy MGH Morphometry Calibration Siemens GE Same Subject Co-registered No Distortion Correction With Distortion Correction Improved across site test-retest reproducibility: alignment of cortical surfaces improves with distortion correction http://spl.harvard.edu Courtesy MGH Function BIRN Calibration Methods for Multi-Site fMRI Study Regional Brain Dysfunction Progression and Treatment of Schizophrenia Human Phantom Trials Common Consortium Protocol 5 Subjects Scanned at All 11 Sites Add'l 15 Controls, 25 Schizophrenics Per Site Statistical Techniques Identify Cross-Site Differences Develop Corrections to Allow Data Pooling Develop Interoperable Post-Processing http://spl.harvard.edu Function BIRN Patients with schizophrenia show abnormal modulation of temporal and frontal regions during semantic processing fMRI response at 3T and 1.5T with identical software and hardware platforms (GE SIGNA) Harvard-MGH Stanford – Lucas Center http://spl.harvard.edu BIRN Query Atlas Morphometry Analysis Converts Raw Image Data into Symbolic form that Allows Complex Queries Can Overlay Population or Subject Data http://spl.harvard.edu Segmentation: MGH; Visualization: BWH BIRN Query Atlas 3D Environment for Neuroscience Queries Identify Studies Drill Down Population Statistics Atlas/Reference Mouse Homology Cellular/Genetic Databases Taxonomy/Homology Gene Expression Protein Localization Web Entrez... http://spl.harvard.edu Real-Time Brain Mapping Concentrate All Computing Power to Treat a Single Patient's Condition Provide Surgical Team with Information-Rich Environment Pre-Operative Images and Analyses Quantied Intra-Operative Changes Integrated Visualization http://spl.harvard.edu Images Provided by Kaus Clinical Problem: Brain Tumors Many Tumor types Glioma, Astrocytoma, etc... Treatment Craniotomy Excision Goal: Preserve Brain Function The “Eloquent” Cortex White Matter Tracts Blood Vessels http://spl.harvard.edu MRT Open Magnet Operating Room Tracked Instruments Image Guided Therapy http://spl.harvard.edu Image Guided Neurosurgery Pre-op Map and Surgical Plan Superimpose on Real Patient Position Use Tracked Instruments to Perform Surgery http://spl.harvard.edu Provided by Leventon et al. Issue: Brain Deformation Centimeter Scale Displacement Pressure Changes when Skull is Opened Surgical Manipulations Retractors Excision “Living” Responses Anesthesiology Inflamation http://spl.harvard.edu Finite Element Method 3D Mesh Generation Intra-Operative Surface Used as Boundary Condition Deformation Field Bring Pre-Op Brain Map to Intra-Op Configuration http://spl.harvard.edu Mesh Generation Initial Uniform Tetrahedral Mesh Clipping and Remeshing Adaptive Refinement http://spl.harvard.edu Material Models Current Software Homogeneous, Linear Elastic 100K Nodes Real Brain is Highly Nonlinear Inhomogeneous, Hyperviscoelastic, Anisotripic Material Properties Change Remember: Patient is Alive! http://spl.harvard.edu Computational Problem Bottleneck is Solving FEM High Resolution Models Accurate Material Models In Clinically-Appropriate Timeframe Iterated Large Sparse Matrix Solutions Numerically the Same as Classical Nonlinear FEM Problems in Engineering Analysis http://spl.harvard.edu Surgical Timeline Before Surgery During Surgery Pre-Operative Mapping Intra-Operative MRI Rigid Registration Tissue Classification Surface Displacement Biomechanical Simulation Visualization Surgical Progress | 0 | 5 | 10 Time (min.) http://spl.harvard.edu | 15 | 20 | 25 | 30 | 35 | 40 | 45 | 50 | 55 | 60 Prediction 1 Grid computing will really take off once users don't have to know they are using it Meaning: don't expect analysis programs will be re-written with grid APIs. Instead, hide grid APIs behind the APIs people already use Example: Freesurfer datasets Over 1G per subject, hundreds of files Dozens dozens of executables and analysis paths Freesurfer implicitly relies on the OS/NFS to manage caching Grid implementation must provide the interface layer Hundreds of programs fit this model http://spl.harvard.edu Prediction 2 Automated Image Segmentation will be the first big success for the Grid in medical imaging Current paradigm is still “one student, one computer, one dataset” Ability to test algorithms on large-N of studies in parallel will qualitatively change the practice Key Challenge: Human-in-the-loop is still required for high quality segmentation http://spl.harvard.edu Prediction 3 Simulation will be the killer app Simulation is needed to integrate our understanding of multiple layers of biology Truly computationally demanding Digital Humans, Virtual Patients, “Body Double”, Holomer... to Allow Tracking of Patient Condition Over Time Model-Based Pathophysiology to Predict Disease Course and Treatment Response Example: New DARPA Virtual Soldier Project U Mich, U Utah, UCSD, U Wash, Stanford, ORNL, Simquest, Mission Research, GE Research and Brigham and Women's http://spl.harvard.edu