Highly Accelerated T_1ρ Imaging in 3 min: Comparison Between Compressed Sensing and Deep Learning Reconstruction

Area of Science:

• Medical Imaging
• Quantitative MRI
• Biomedical Engineering

Background:

• Quantitative T1ρ mapping of knee cartilage is crucial for diagnosing joint diseases.
• Traditional T1ρ mapping requires long scan times, limiting clinical application.
• Accelerated imaging techniques like compressed sensing (CS) and deep learning (DL) aim to reduce scan duration.

Purpose of the Study:

• To compare the performance of CS and DL for accelerated T1ρ mapping in knee cartilage.
• To evaluate both retrospectively and prospectively undersampled reconstruction methods.
• To assess the accuracy and repeatability of DL versus CS in knee cartilage imaging.

Main Methods:

• T1ρ-weighted 3D MAPSS sequence on a 3T MRI scanner was used for T1ρ map generation.
• Retrospective undersampling involved two schemes (UF4_4echo, UF8_8echo); prospective undersampling was also performed.
• Reconstructions were compared against reference scans using median normalized absolute differences (MNADs), concordance correlation coefficient (CCC), and coefficient of variation (CV).

Main Results:

• For retrospective undersampling, both CS and DL achieved high CCC values, with DL showing slightly lower MNADs.
• For prospective undersampling, DL reconstructions demonstrated significantly higher CCC and lower CV compared to CS.
• Scan-rescan CV for DL prospective reconstructions (UF4_4echo: 2.4%, UF8_8echo: 2.8%) were comparable to the reference (2.57%) and better than CS.

Conclusions:

• Deep learning (DL) reconstruction offers superior performance compared to compressed sensing (CS) for prospectively undersampled T1ρ mapping in knee cartilage.
• DL-based acceleration holds significant promise for reducing scan times in quantitative knee MRI without compromising diagnostic quality.
• The findings support the clinical utility of DL for faster and more efficient knee cartilage assessment.