Novel MOS prediction models for compressed medical image quality

Area of Science:

• Medical Imaging
• Image Processing
• Data Compression

Background:

• Assessing diagnostic information in compressed medical images is crucial.
• Current methods for quality assessment can be time-consuming and subjective.

Purpose of the Study:

• To develop novel models for predicting the quality of compressed medical images.
• To establish upper limits for image compression thresholds while preserving diagnostic information.

Main Methods:

• Developed models correlating objective quality metrics (PSNR, SSIM) with subjective Mean Opinion Scores (MOS).
• Validated models using JPEG, JPEG2000, and SPIHT compression of CT scan images.
• Compared model-predicted MOS with experimental MOS from six human observers.

Main Results:

• High correlation (≥0.87) found between model-generated and experimental MOS.
• Prediction accuracy for Percentage Ріtеrеnсе Difference (PRD) was ≤13% within the 0.05-2.0 bits/pixel compression range.
• Models demonstrated observer-independent MOS prediction capabilities.

Conclusions:

• The developed models offer a reliable method for observer-independent quality assessment of compressed medical images.
• These models can help determine optimal compression thresholds, preserving diagnostic relevance.
• The approach is suitable for teleradiology, reducing the need for extensive experimental MOS evaluations.