Predicting human detection performance in magnetic resonance imaging (MRI) with total variation and wavelet sparsity...

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Magnetic resonance imaging (MRI) reconstruction often uses total variation (TV) and wavelet sparsity regularization to reduce noise and undersampling artifacts.
• While effective, these methods can introduce their own unique artifacts, impacting image quality and diagnostic accuracy.

Purpose of the Study:

• To extend human observer performance modeling to wavelet regularization and combined wavelet-TV regularization in MRI.
• To evaluate the impact of TV and wavelet regularization on the detection of small lesions in undersampled FLAIR brain images.

Main Methods:

• Small lesions were inserted into undersampled (acceleration factor 3.48) fastMRI FLAIR brain data.
• Images were reconstructed using a range of regularization parameters for TV and wavelet methods.
• A sparse difference-of-Gaussians (S-DOG) model observer was used, with its noise level calibrated to human performance in two-alternative forced choice (2-AFC) studies.

Main Results:

• The S-DOG model generally matched human observer performance across regularization parameters.
• At high regularization parameter values, the S-DOG model outperformed average human observers.
• Neither TV nor wavelet regularization, individually or combined, enhanced human observer performance for lesion detection in this task.

Conclusions:

• Current regularization techniques in MRI reconstruction do not necessarily improve human detection of subtle lesions.
• Further research is needed to develop regularization methods that better align with human visual perception and improve diagnostic tasks.