The "hidden noise" problem in MR image reconstruction.

Area of Science:

• Medical Imaging
• Image Reconstruction
• Quantitative Analysis

Background:

• Quantitative error metrics like root mean squared-error and structural similarity index are standard for evaluating image reconstruction.
• These metrics compare reconstructed images to fully sampled reference data, assuming the reference is a perfect gold standard.

Purpose of the Study:

• To investigate the impact of "hidden noise" in reference data on the performance ranking of image reconstruction methods.
• To determine if conventional error metrics are confounded by noise in reference data.

Main Methods:

• Utilized experimental and simulated k-space data across various image reconstruction techniques.
• Compared performance metrics derived from noisy reference data versus higher-quality reference data.

Main Results:

• Reconstruction methods ranked highest with noisy reference data differed significantly from those ranked highest with higher-quality data.
• Conventional metrics led to suboptimal reconstruction choices when evaluated against noisy references.
• Alternative error metrics that account for noise reduced these discrepancies.

Conclusions:

• "Hidden noise" in reference data can significantly mislead the evaluation and ranking of image reconstruction methods using conventional metrics.
• Employing alternative error metrics that better handle noise is crucial for accurate performance assessment in image reconstruction.