Comparison of model-based versus deep learning-based image reconstruction for thin-slice T2-weighted spin-echo...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Model-based iterative reconstruction (MBIR) is a standard technique for MRI image reconstruction.
• Deep learning (DL) based methods offer potential for improved image quality in MRI.

Purpose of the Study:

• To compare a novel deep learning (DL) reconstruction with a previous model-based iterative reconstruction (MBIR).
• To evaluate the signal-to-noise ratio (SNR) improvement in high-resolution (1 mm) T2-weighted spin-echo (T2SE) prostate MRI.

Main Methods:

• Quantitative phantom studies assessed contrast and spatial resolution.
• Radiological evaluation in 17 subjects undergoing prostate MRI at 3.0 Tesla.
• Comparison of MBIR and three DL reconstruction levels (Low, Medium, High) on thin-slice T2SE images.

Main Results:

• All DL reconstruction levels demonstrated improved SNR compared to MBIR.
• DL Low and Medium levels were superior to MBIR in most evaluation criteria.
• DL Low and Medium showed better contrast fidelity than DL High, with DL Medium preferred in 44/51 evaluations.

Conclusions:

• Deep learning reconstruction significantly enhances SNR in thin-slice T2SE prostate MRI while preserving contrast.
• Overly aggressive DL enhancement (DL High) can compromise radiological sharpness and contrast fidelity.