Compression of Real-Time Cardiac MRI Video Sequences

EE 368B Final Project

December 8, 2000

Neal K. Bangerter and Julie C. Sabataitis

Overview• Real-time cardiac MRI imaging

– New technology

– 128 x 128 pixels, 18 frames / sec

• Compression of cardiac sequences for remote diagnosis:– Motivation

– What PSNR is necessary to preserve diagnostic utility of sequences?

– What compression techniques work best on these real-time cardiac sequences?

– What channel bit-rate is required for streaming of these sequences?

Project Goals

• Implement video compression algorithm that supports:– Frame-difference encoding– Motion Compensated Prediction (MCP)– Long-term memory MCP

• Optimize MCP parameters for real-time cardiac MRI studies

• Determine acceptable PSNR for diagnosis

• Identify compression technique which yields lowest bit-rate at determined PSNR

• Wiegand, Zhang, Girod (1997): decrease prediction error by increasing block matching to search many previous frames

• Bit savings from better prediction should be larger than number of bits needed to send displacements (dx, dy, dt)

• MCP Parameters:– Block size

– Search range: maximum absolute value of dx, dy

– Frame buffer size: number of previous frames used for comparison

MCP with Long-Term Memory

Initial Exploration of MCP on Original Sequences using Matlab

• MCP (long-term and single-frame) with uniform quantization of DCT coeff.

• Smaller displacement vectors for single-frame MCP, similar error images for both

• Block indices for time buffer frame selected was often previous frame– Suggests strong frame-to-

frame correlation

Displacement vectors

Long-term MCP Single-frame MCP

Mesh plots of error images

Exploration of Matlab MCP on Synthetic Periodic Sequence

• Five frames of short-axis study repeated

• Expect three things of long-term MCP:– Time buffer indices

should be 5 at each block– Displacement vectors

should be 0– Error image should

consist of only quantization noise

Displacement vectors

Long-term MCP Single-frame MCP

Mesh plots of error images

Matlab MCP on Temporally Sub-Sampled Sequences

• 2/3 of image data shared between successive frames

• Sampled sequences temporally to remove dependencies:– No data shared: 6 fps

– 1/6 of data shared: 9 fps

Displacement vectors

Long-term MCP Single-frame MCP

Mesh plots of error images

C Implementation Features

• Variable block size, search range, and frame buffer size

• Zig-zag and run-level encoding of 8x8 DCT blocks

• Lagrangian cost function using block MSE and bit cost of motion vectors (dx, dy, dt)

Testing

• Periodic video sequence: 10 frames repeated

• PSNR of predicted image should increase significantly beyond 11th frame

• MCP with buffer >= 10 frames should yield significant compression gains

Optimizing MCP Parameters

• Try 35 different MCP parameter combinations:– 16x16, 8x8, and 4x4 block size

– 2, 4, and 8 pixel search range

– 1, 2, 4, 8, and 16 frame buffer size

• Run each at 7 different quantization levels to generate 35 PSNR curves

• Frame-difference and intra-frame PSNR curves also generated

• High PSNR• Long-term MCP• 4x4 blocks• 4 pixel search

range• 16 frame buffer

• Low PSNR• Frame-

difference coding best

Optimization Results

Determination of Acceptable PSNR

• Presented videos at different PSNR to cardiologist

• 30 to 31 dB sufficient for current applications (wall motion assessment, coronary imaging)

• Very few cardiologists familiar with cardiac MRI

• New technology: as quality increases, new applications will emerge that may have different PSNR requirements

Conclusions• Current applications require PSNR of 30-31 dB to preserve

diagnostic utility

• At this PSNR, simple frame-difference coding yields best compression– Original 2.3 Mbps

– Compressed ~70 Kbps

• Current real-time cardiac MRI video experiences little to no gain in PSNR at a given bit-rate (generally < 1 dB) when using long-term memory MCP vs. frame-difference encoding– Strong frame to frame correlation

– Limited motion often confined to a small portion of the image

Future Work

• Capabilities of real-time MRI likely to increase– Revisit MCP techniques as images become less noisy

and have higher resolution

• Development of metrics for evaluation of “acceptable” image distortion levels for various kinds of diagnostic studies

• Integration of video-compression techniques with remote-diagnosis systems

• Compression of spatial frequency MRI data prior to reconstruction

Acknowledgements

• Markus Flierl for zig-zag DCT compression code and for his help whenever we showed up at his office

• Authors of the CIDS library of C functions for image processing and compression

• Bob Hu for evaluation of real-time sequences at various PSNR levels

• Krishna Nayak for providing real-time cardiac MRI sequences