Modeling human observer detection for varying data acquisition in undersampled MRI for two-alternative forced choice...

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging (MRI)
• Image Reconstruction

Background:

• Undersampling in k-space accelerates MRI data acquisition.
• This study investigates image quality and human perception under varying degrees of undersampling.
• Reconstruction algorithms like multi-coil SENSE are crucial for recovering images from incomplete data.

Purpose of the Study:

• To evaluate human observer performance in MRI tasks with 1D undersampled k-space data.
• To assess the impact of aliasing versus blurring on image perception.
• To model human performance using computational observer models.

Main Methods:

• Acquired 20% of k-space data with a 5x undersampling factor.
• Varied low k-space frequency content from 0% (aliasing) to 20% (blurring).
• Used 2-alternative forced choice (2-AFC) and forced localization tasks with human observers.
• Modeled observer performance using sparse-difference of Gaussians (SDOG) Hotelling observer models.

Main Results:

• Human observer performance in 2-AFC tasks was consistent across conditions.
• Forced localization performance improved with minimal blurring (2.5%) but degraded in pure aliasing (0%).
• Both symmetric and asymmetric SDOG models predicted 2-AFC performance, with the symmetric model showing a slight advantage.
• A symmetric 4-channel SDOG model accurately modeled forced localization results.

Conclusions:

• Human perception of image quality in undersampled MRI is robust to varying aliasing/blurring levels, except in pure aliasing.
• Computational observer models, particularly the symmetric SDOG model, can effectively predict human performance in these tasks.
• These findings support the feasibility of accelerated MRI acquisition through k-space undersampling while maintaining diagnostic utility.