Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Medical Image Processing on NVIDIA TK1/TX1

900 views

Published on

Che-Lun Hung

Published in: Technology

Medical Image Processing on NVIDIA TK1/TX1

  1. 1. TAIPEI | SEP. 21-22, 2016 Che-Lun Hung, 2016/9/21 MEDICAL IMAGE PROCESSING ON NVIDIA TK1/TX1
  2. 2. 2 AGENDA Medical Image Brain MRI Image Segmentation Fuzzy C-Means Method on TK1/TX1 Genetic Fuzzy C-Means Method on Multiple TK1s
  3. 3. 3 MEDICAL IMAGE
  4. 4. 4 MEDICAL IMAGE X-Ray Ultrasound Image resource : wiki CT MRI Image Source: https://en.wikipedia.org/
  5. 5. 5 WHY NEED MEDICAL IMAGE PROCESSING? 9/26/16 Computer-aided diagnosis Description of Lension Image Source: http://www.dailyherald.com/ Image Source: http://artificialintelligencefordummies.weebly.com/
  6. 6. 6 TK1/TX1 TK1 Kepler 192 CUDA cores 2G Memory TX1 Maxwell 256 CUDA cores 4G Memory
  7. 7. 7 BRAIN MRI IMAGE SEGMENTATION
  8. 8. 8 BRAIN MRI IMAGE Image Source: http://brainweb.bic.mni.mcgill.ca/brainweb/
  9. 9. 9 SEGMENTATION Image Source: http://brainweb.bic.mni.mcgill.ca/brainweb/
  10. 10. 10 CLUSTERING METHODOLOGY
  11. 11. 11 SEGMENTATION FOR BRAIN MRI
  12. 12. 12 FUZZY C-MEANS METHOD ON TK1/TX1
  13. 13. 13 FUZZY C-MEANS METHOD Jm = uij n d xi,θj( ) j=1 c ∑ i=1 n ∑
  14. 14. 14 EXPERIMENTAL PLATFORMS TK1/TX1 The host (CPU) Intel Xeon E3-1231 v3 3.40GHz with 64GB RAM Data Size 1150*1280 9/26/16
  15. 15. 15 SEGMENTATION RESULTS Image Source: http://brainweb.bic.mni.mcgill.ca/brainweb/
  16. 16. 16 PERFORMANCE 1 7.2 12.9 0 2 4 6 8 10 12 14 Intel E3 CPU TK1 GPU TX1 Speedup Platform IMG Size= 1150*1280
  17. 17. 17 DIFFERENT PRECISION FORMAT 5.93 2.22 3.81 1.42 0 1 2 3 4 5 6 7 TK1 GPU double TK1 GPU float TX1 GPU double TX1 GPU float Time(seconds) Platform
  18. 18. 18 TK1/TX1 MAXIMUM PERFORMANCE MODE 5.93 2.22 3.81 1.42 5.45 2.15 3.69 1.39 0 1 2 3 4 5 6 7 TK1 GPU double TK1 GPU float TX1 GPU double TX1 GPU float Time(seconds) Platform Kernel time(Normal) Kernel time(HP)
  19. 19. 19 DIFFERENT MEMORY COPY MODES 2.15 1.39 2.38 1.5 2.37 1.45 0 0.5 1 1.5 2 2.5 TK1 GPU float TX1 GPU float Time(s) Platform Kernel time Kernel time(Zero-copy) Kernel time(Unified memory)
  20. 20. 20 GENETIC FUZZY C-MEANS METHOD ON MULTIPLE TK1S
  21. 21. 21 GENETIC FUZZY C-MEANS METHOD Image Source: http://http://www.turingfinance.com/
  22. 22. 22 MESSAGE PASSING INTERFACE (MPI) Image Source: https://computing.llnl.gov/tutorials/mpi/
  23. 23. 23 MULTIPLE GPU MPI+CUDA
  24. 24. 24 EXPERIMENTAL PLATFORMS TK1/TX1 The hosts (CPU) Intel Xeon E5-2620 2.5GHz, E3-1231 V3 3.4GHz, and I7-4280K 3.7GHz. Data Size 1150*1280 9/26/16
  25. 25. 25 DIFFERENT POPULATION SIZES (a) (b) (c) (d) (e) (f) (a) population size = 30, (b) population size = 60, (c) population size = 90, (d) population size = 120, (e) population size = 150, (f) population size = 180. Image Source: http://brainweb.bic.mni.mcgill.ca/brainweb/
  26. 26. 26 PERFORMANCE
  27. 27. 27 SUITABLE FOR ALL MEDICAL IMAGE PROCESSING TECHNOLOGIES?
  28. 28. 28 MAMMOGRAPHIC IMAGE
  29. 29. 29 TEXTURE-BASED IMAGE PROCESSING (1)
  30. 30. 30 TEXTURE-BASED IMAGE PROCESSING (2) 127*127127*127127*127127*127 ……………127*127 127 * 127 threads
  31. 31. 31 EXPERIMENTAL PLATFORMS TK1/TX1 The host (CPU) Intel Xeon E3-1231 v3 3.40GHz with 64GB RAM Data Size 1123*1751 9/26/16
  32. 32. 32 PERFORMANCE 0 5000 10000 15000 20000 25000 30000 35000 Matlab with NVIDIA GTX 980 Matlab with Intel Xeon E3-1231 NVIDIA TX1 NVIDIA TK1 550 1752 20260 31469 Executiontime(s) Device
  33. 33. 33 MORE POSSIBILITIES FOR GPU ON MEDICAL IMAGE PROCESSING
  34. 34. TAIPEI | SEP. 21-22, 2016 THANK YOU

×