Video Compression for Medical Imaging

1
Video Compression for Medical Imaging

• by David Gibson

2
Contents

• Part 1 Compression Background
• Fundamentals of Compression
• Video Motion Compensation
• Part 2 Medical Imaging
• Example of the data JPEG /Wavelet encoding
• Motion compensation
• Region of interest (ROI) coding

3
Part 1

• Video Compression Review

4
Foundations of Compression

• The Foundations of Compression involves looking
  at the data.

5
Foundations of Compression
6
Foundations of Compression
DCT
7
Video Compression
8
Main Classification of Video Compression Methods

• Intra-frame methods
• Uses single frames
• e.g. MJPEG - JPEG applied to video

• Inter-frame methods
• Uses temporal information
• e.g. MPEG-1/2, H.263
• Usual approach to video compression

9
Inter-frame methods

• Use Motion Compensation

10
Motion Compensation

• Exploitation of temporal redundancy.

Frame 30
Frame 31
Motion Compensation
11
How Do We Motion Compensate?

• Compensate each pixel separately with its own
  motion vector?

Motion Data
Error Data

• Huge amount of motion data - More data than the
  original image!
• Cant afford to motion compensate each individual
  pixel.

12
Solution

• One motion vector for a group of pixels.
• Based on looking at the data.

13
Block Matching

• Foundation of most current video coders (MPEG
  1/2, H.261/3).

14
Conclusions (part 1)

• Presented a brief summary of video compression
  methods

15
Part 2

• Video Compression of Medical Images

16
Medical Imaging

• Angiogram Video
• Pictures taken of the heart at 30 frames/second
• 512x512 images - 8 bits/pixel
• Typical procedure - 5 minutes
• Resulting in 2.5GBytes of data per patient.
• _at_64Kbits/sec - 80 hours.
• _at_10Mb/sec - 30 minutes.

17
Summary

• Going to look at 3 aspects of the research weve
  been doing
• Example of the data JPEG/Wavelet encoding
• Motion compensation
• Region of interest (ROI) coding

18
Example Angiogram Sequence
19
Example JPEG Coding
20
Still Frame Coding Methods Wavelet

• Similar frequency approach to DCT.
• But considered to give better results.
• Operation on the whole image.

21
JPEG/Wavelet Comparison
22
Use an off the shelf video coder?

• Typical results for an angiogram image _at_0.8bpp.
• Comparison of intra- and inter-frame methods
  using DCT.

• Motion compensation performs badly for this type
  of data.
• Key Point Compression effectiveness depends upon
  the data

23
Motion Compensation - Failure?

• Conventional motion compensation assumptions
• Distinct, opaque objects moving simply.
• Also, angiogram images contain high frequency
  uncorrelated texture.

24
Motion Compensation - Failure?

• Objects in angiograms are partially transparent.
• Image is made up of several layers of bones and
  tissue, all moving differently.
• Conventional motion compensation model doesnt
  apply well.

25
Region of Interest (ROI) Coder

• Aim is to shift the allocation of bits from
  uninteresting areas of the image to more
  interesting ones.
• Makes more efficient use of the available bits.

26
ROI Example Simple Case

• Manual segmentation.

ROI
non-ROI
27
Example ROI coder

• Example of transferring bits from non-ROI to ROI

28
ROI Simple Case - Results
RD Graph with ROI - DFD Data (Global MC - M.Black)
8
No ROI (baseline comparison)
ROI Distortion
7
Non-ROI Distortion
6
5
Distortion (RMS error)
4
3
2
1
0
0
0.5
1
1.5
2
2.5
3
3.5
4
Rate (bits/pixel)

• Much lower error in the ROI at the expense of the
  non-ROI.

29
Key Aim

• Reallocate bits from diagnostically unimportant
  areas into diagnostically interesting ones

30
Eye Tracking (proof of concept)

• Experiment to identify key areas of an angiogram
  image.

31
Example Results (Expert)
32
Example Results (Sandra)
33
Eye Tracking

• Significant areas of the image are not directly
  examined.

34
Quality Measure and Results

• Methods of measuring image quality
• Classical RMS - Measure of intensity level
  difference for each pixel.
• Perceptual measure - Takes in to account the
  observer.

35
Quality Measure and Results

• Perceptual measurement of image quality.

Poor
Perfect
1
2
3
4
5
Original
Compressed
36
Whats next for video compression research?

• More efficient compression methods - to better
  take advantage of data (e.g. object based)
• Perceptual coding - introducing the viewer into
  the equation

37
Questions?