Using deep feature distances for evaluating the perceptual quality of MR image reconstructions

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Traditional image quality (IQ) metrics like PSNR and SSIM often fail to align with expert radiologist assessments of diagnostic IQ in MR image reconstruction.
• Developing novel metrics is crucial for accurately evaluating the perceptual quality of accelerated MR image reconstructions.

Purpose of the Study:

• To develop and evaluate deep feature distances (DFDs) as improved perceptual IQ metrics for MR image reconstruction.
• To assess the impact of domain shifts on DFD performance.
• To compare DFDs against established and state-of-the-art IQ metrics.

Main Methods:

• Compared standard IQ metrics (PSNR, SSIM) with DFDs (in-domain, domain-adjacent, out-of-domain) and other metrics (VIF, NQM, HFEN).
• Evaluated metric performance by correlating with expert radiologist scores for accelerated MR image reconstructions.
• Assessed metric sensitivity to acquisition noise and distortions.

Main Results:

• All DFDs and HFEN demonstrated stronger correlations with radiologist-perceived diagnostic IQ than PSNR and SSIM.
• The performance of DFDs was comparable to radiologist inter-reader variability.
• Out-of-domain DFDs performed as well as in-domain and domain-adjacent DFDs.

Conclusions:

• DFDs and HFEN offer a more holistic evaluation of MR image reconstruction perceptual quality when used with traditional metrics.
• General vision encoders can effectively assess visual IQ for MR images, even without specific MR training data.