1 / 58

Handling large biological datasets Part 1

Handling large biological datasets Part 1. PSC Volume Browser performance issues and mouse heart analysis with the Duke CIVM. Initial PSC-VB Development during U.Michigan/NLM project. Client-Server System Architecture. vs.psc.edu is located in PSC’s machine room. www.psc.edu.

wylie-keith
Download Presentation

Handling large biological datasets Part 1

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Handling large biological datasetsPart 1 PSC Volume Browser performance issues and mouse heart analysis with the Duke CIVM

  2. Initial PSC-VB Development during U.Michigan/NLM project

  3. Client-Server System Architecture vs.psc.edu is located in PSC’s machine room. www.psc.edu

  4. Implementation Issues/Design • Problems (distributing CDs to PCs): • Memory limitations • 32 bit chips only address ~2 GB of physical memory • Avoid disk seeks • IDE Disks have slow disk seeks which results in direction-dependent speeds • Maintenance of multiple image datasets with corresponding annotation • How to deal with a work in progress? • Tie into educational material • Network bottlenecks • Last mile problem, network aware…

  5. Implementation Issues/Design Solution (Network-bound Client-Server): • Use memory-based server representations • 64 bit architecture, fast memory access • amortize additional costs across all users • Apply volumetric compression and progressive resolution techniques • Take advantage of specialized client hardware • Graphics (thank the gaming industry) • Leverage network tuning from PSC’s Web100

  6. Relevant Technology Evolution • Storage • Disk (~4 GBytes/$, 1 TByte/drive, ~100 MB/sec/drive, ~100-200 accesses/sec • Main memory (~$40/GByte, 1-8 GBytes/CPU) • Solid state drives – very high cost • Flash drives (~1/2 cost of DRAM, 500us access, 30 MB/s bandwidth) • Computation – Moore’s law still holds • Networking – improving rapidly but by jumps • Graphics hardware – effective rates improving faster than Moore’s law and now useful for general computation but still memory limited.

  7. Methodology items of note: • Borrowed lossless H.264 transform • Used a cooperative development strategy • Communicate release instructions on web • Evaluation team provided guidance user question forms/interviews • Demonstrated CIVM database linkage by using PSC-VB as a helper application • Tried to maintain broad applicability

  8. Network Parameter Tuning

  9. Major NLR routes

  10. The PSC - CIVM pathway • NLR is the main pathway • Provides state-of-the-art performance • RTTs have dropped from ~30 ms to 15 ms • Jitter is now < 10 ms • Center to center can reach 200 Mbits/sec • Desktop end to end > 30 Mbits/sec

  11. 49 50 51 52 33 34 35 36 52 17 18 19 20 36 1 2 3 4 20 56 1 2 3 4 4 40 24 60 5 6 7 8 8 44 28 64 9 10 11 12 12 48 32 1 seek x 16 cubes 4 seeks x 4 cubes 16 seeks x 1 cubes 13 14 15 16 16 4 20 36 52 1 2 3 4 1 2 3 4 8 24 40 56 5 6 7 8 17 18 19 20 12 28 44 60 9 10 11 12 33 34 35 36 16 32 48 64 13 14 15 16 49 50 51 52 Impact of Disk Flat vs BlockDisk Data Format

  12. Disk transfer vs access time

  13. Optimal Tube Sizes

  14. MMAP vs READ speed

  15. Disk access time is more critical than network latency Cut plane FOM vs record size Multi user elevator ordering effects. Single access vs seek distance.

  16. Hierarchical Wavelet Concept

  17. Hierarchy of Wavelet Scales

  18. 1150:1 205:1 85:1 32:1 visually lossless VM Film Hierarchical Wavelet Organization Eighth Resolution Add detail Quarter Resolution Add detail Half Resolution Add detail ~2:1 fully lossless Full Resolution

  19. Hierarchical Normal Mesh Compression Produced by Jason McAllister

  20. Adapting PSC-VB for mouse heart 4D volume measurement with Duke CIVM

  21. CIVM Large Buckets of Data • Very Large 3D Arrays- MR Histology: (1k x 1k x 4k) • 4D Arrays: (3D + time) - cardiac micro-CT or MRM - perfusion micro-CT or MRM • Multimodality Data- combined micro-CT/DSA - combined micro-CT and micro-PET

  22. Leading Edge Micro MRI

  23. Results: Pre-natal Development SE 3D sequence TR / TE = 75 / 5.23 ms Matrix = 1024 x 512 x 512 FOV = 20 x 10 x 10 mm Nex = 2 Tacq = 6 h 22 min

  24. Reconfigurable Micro CT

  25. Mouse Project goals: • Provide technology for serving and viewing large 4D datasets • Provide secure data installation and access • Provide networked tools for 4D data visualization and analysis • Provide I2 accessible online data repository • Evaluate the effectiveness of resulting tools and techniques for usability, effectiveness and applicability to other areas.

  26. Triggered Acquisition • -X-ray : 80 Kvp, 150 mA and 10ms(flux sufficient to fill the detector wells to ~ 25 %). Exposure 60 mR/proj • Projections acquired on1900 with a step angle of 0.750. • Scanning time about 8-10 mins

  27. 4D Cardiac Micro-CT in Mice 100x100x100 microns x 10 ms

  28. Conventional Analysis of Cardiac Function Systole: R+ 30 ms Table1: Cardiac function estimation in (n=5) mice using the micro CT

  29. Live Mouse Micro CT Goals • Improve accuracy of LV measurement • Reduce contrast agent dose • Reduce radiation exposure • Reduce analysis time • Reduce manual analysis intervention • Enable time studies of individual animals

  30. MI: Hyperenhancement 5 days 35 days Nahrendorf M, Badea C, Hedlund LW, Figueiredo JL, Sosnovik DE, Johnson GA, Weissleder R. High Resolution Imaging of Murine Myocardial Infarction With Delayed Enhancement Cine Micro-CT. Am J Physiol Heart Circ Physiol. 2007 Micro-CT Study of Myocardial Infarction (MI) in Mice • MI mouse model by LAD ligation • Scanned at 5 days and 5 weeks post MI! • Goal: MI size and cardiac function

  31. Wall Motion Analysis Cardiowarp S/w

  32. An alternative approach to LV volume measurement • Take advantage of known binary mixture model • Avoid difficulties of segmentation methods • Account for unresolved detail and motion • Tolerate high noise levels and artifacts • Provide numerical error estimates • Trade SNR against resolution - C. Shannon 1948 • C = W log ((S+N)/N) • Use all ROI data to form a simple ratio measurement • Use measured result to constrain segmentations

  33. Cardiac Function : MI vs. Controls

  34. MLP null C57BL/6 MICRO-CT FORMORPHOLOGICAL AND FUNCTIONAL PHENOTYPING OF MLP NULL MICE

  35. Live MicroCT vs. Fixed Micro MR

  36. Mammalian Heart Anatomy

  37. Nested ROI and targeted blood/muscle sampling

  38. PSC-VB Cardiac Ejection

  39. Effects of Contrast Agent and # of CT projections 0.125 ml Contrast Agent 0.5 ml Contrast Agent 63 projections 380 projections 63 projections 380 projections

  40. Effects of filtering and # projections over full ROIat high contrast dose

  41. Effects of filtering and # projections over full ROIat low contrast dose

  42. Compute volume directly from gray values FractBlood = (AvgROI – AvgMuscle) / (AvgBlood – AvgMuscle) The resulting error is small despite high voxel noise. The ROI volume is exactly known from its construction. VolumeBlood = FractBlood * VolumeROI

  43. Approach is similar to an analog to digital converter

  44. Histograms of isolated blood and muscle vs. contrast agent and #projections 0.125 ml contrast agent 0.5 ml contrast agent Black histograms are from 380 projections. Gray histograms are from 63 projections.

More Related