Modeling human observer detection in undersampled magnetic resonance imaging reconstruction with total variation and...

Area of Science:

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

Background:

• Undersampled MRI requires regularization to reconstruct images.
• Task-based image quality assessment evaluates how reconstruction affects diagnostic performance.
• Total Variation (TV) and wavelet regularization are common techniques.

Purpose of the Study:

• To evaluate the impact of TV and wavelet regularization on human detection performance in undersampled MRI.
• To validate a model observer's ability to predict human performance in these tasks.

Main Methods:

• Human observer studies using two-alternative forced choice (2-AFC) on undersampled MRI data.
• Reconstruction utilized TV and wavelet sparsity constraints at a 3.48 undersampling factor.
• A Sparse Difference-of-Gaussians (S-DOG) model observer was employed and validated against human data.

Main Results:

• Human detection performance remained stable across a range of regularization parameters before decreasing at higher values.
• The S-DOG model observer reasonably predicted human performance for both TV and wavelet regularization.
• The model observer slightly overestimated performance at high regularization levels.

Conclusions:

• The S-DOG model observer is effective for predicting human performance in task-based assessments of undersampled MRI with TV and wavelet regularization.
• Task performance in undersampled MRI is robust to moderate regularization but degrades with excessive application.