Mutual Information Correlation with Human Vision in Medical Image Compression

Area of Science:

• Medical Imaging
• Image Processing
• Signal Compression

Background:

• Lossy compression algorithms impact image quality variably across contrast regions.
• Low contrast areas experience greater quality decline than high contrast regions at equal compression ratios.
• Objective image quality metrics require closer alignment with human visual perception.

Purpose of the Study:

• To evaluate objective image quality metrics for assessing lossy compression effects on different contrast areas.
• To determine if metrics like Peak Signal-to-Noise Ratio (PSNR) and Mutual Information (MI) correlate with human visual perception.
• To develop improved metrics that account for contrast variations in medical images.

Main Methods:

• Measured PSNR and MI for discrimination between contrast areas in a SMPTE electronic pattern under lossy compression.
• Applied measurements to compressed medical images (CT, MR, CRX) from different contrast regions.
• Developed weighted PSNR (W-PSNR) and weighted MI (W-MI) to account for gray value and contrast.

Main Results:

• MI correlated with human vision in SMPTE patterns, CT, and MR, but not CRX.
• PSNR did not consistently align with human vision results.
• Both W-PSNR and W-MI demonstrated responsiveness to gray values and contrast areas for quality estimation.

Conclusions:

• W-PSNR and W-MI effectively discriminate between contrast areas based on compression ratios.
• W-MI showed superior performance compared to W-PSNR in reflecting human visual perception.
• W-MI is proposed as a reliable image quality index for lossy compressed medical images.