Super-resolution musculoskeletal MRI using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Deep Learning for MRI

Background:

• Generating thin-slice knee MRI is crucial for detailed diagnosis.
• Current interpolation methods can degrade image quality.
• Super-resolution techniques offer potential for improving MRI resolution.

Purpose of the Study:

• To develop a super-resolution technique using convolutional neural networks (CNNs) for generating thin-slice knee MRI from thicker slices.
• To compare the performance of this CNN-based method against traditional interpolation techniques.

Main Methods:

• A 3D CNN, DeepResolve, was implemented to learn residual transformations between high- and low-resolution MRI slices.
• The network was trained on 124 double echo in steady-state (DESS) datasets and tested on 17 patients.
• Quantitative metrics (structural similarity, peak SNR, RMS error) and qualitative reader studies were used for comparison against tricubic interpolation, Fourier interpolation, and sparse-coding super-resolution.

Main Results:

• DeepResolve significantly outperformed tricubic, Fourier, and sparse-coding super-resolution methods in quantitative image quality metrics.
• Reader studies showed DeepResolve significantly outperformed tricubic interpolation in all assessed image quality categories and overall diagnostic quality.
• High inter-reader reliability (κ = 0.73) was achieved in the qualitative assessment.

Conclusions:

• DeepResolve effectively generates high-resolution thin-slice knee MRI from lower-resolution thicker slices.
• The developed CNN-based super-resolution method achieves superior quantitative and qualitative diagnostic performance compared to conventional and state-of-the-art techniques.