Interactive Exploration of Large Remote Micro-CT Scans - PowerPoint PPT Presentation

About This Presentation
Title:

Interactive Exploration of Large Remote Micro-CT Scans

Description:

Large medical datasets (10 100 GB) Centrally (remotely) stored image data. Interactive exploration. Design Overview. Goal: efficient external data access ... – PowerPoint PPT presentation

Number of Views:60
Avg rating:3.0/5.0
Slides: 10
Provided by: markusg1
Learn more at: https://web.eecs.utk.edu
Category:

less

Transcript and Presenter's Notes

Title: Interactive Exploration of Large Remote Micro-CT Scans


1
Interactive Exploration of Large Remote Micro-CT
Scans
  • Prohaska, Hutanu, Kähler, Hege (Zuse Institut
    Berlin)

2
Motivation
  • Large medical datasets (10 100 GB)
  • Centrally (remotely) stored image data
  • Interactive exploration

3
Design Overview
  • Goal efficient external data access
  • Interface to regular grid data
  • multidimensional array of data elements
  • subblocks at various resolutions can be accessed.
  • Asynchronous access using a background thread
  • Under the hood HDF5 as data container

4
HDF5 and Data Layout
  • HDF5 supports
  • subblock/ hyperslab selection
  • external I/O drivers
  • only synchronous I/O ?
  • Data is stored at highest resolution using a
    chunked layout (blocksize 256KB 2MB)
  • In addition, a low-res preview version of data
    is generated (adds 15 of data)

5
HDF5 and Remote Data Access
  • When handling remote data, the external I/O
    driver should know which high-level function it
    serves (virtual data access layer).
  • GridFTP was chosen as underlying client-server
    data access architecture.

6
Rendering
  • Hierarchical multi-resolution approach
  • Spherical ROI visualized in high-res, other data
    displayed at coarser resolution
  • Octree structure used as multilevel
    representation of data volume, to find blocks
    intersecting ROI.
  • 3D texture based volume rendering (octree based
    back-to-front rendering)

7
Results
  • Measurements
  • Responsiveness (time to load one block) latency
    block size/ bandwidth
  • Throughput
  • Data server dual Xeon 1.7GHz with 170GB storage
    (at 15MB/s read rate)
  • Network latency 0.2ms (LAN), 360ms (WAN)

8
Results (Contd.)
  • Dataset 2048 x 2048 x 1024, 2 bytes per voxel,
    8GB of data.
  • Preprocessing (generation of preview version)
    takes 25 minutes and adds 2GB of extra data.
  • Block sizes used 128 x 128 x 64 (2MB), 64 x 64 x
    32 (256KB)

9
Actual Results
  • Responsiveness
  • 2MB block 1.9s/blk (WAN), 0.7s/blk (LAN)
  • 256K block 1.4s/blk (WAN), 0.1s/blk (LAN)
  • Throughput
  • 2MB block 8.3 MBit/s (WAN), 23MBit/s (LAN)
  • 256K block 1.4MBit/s (WAN), 20MBit/s (LAN)
  • Basically, as expected. Trade off between
    responsiveness and throughput.
Write a Comment
User Comments (0)
About PowerShow.com