Deep Learning-Based Accelerated MR Cholangiopancreatography Without Fully-Sampled Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Magnetic Resonance Imaging

Background:

• Magnetic Resonance Cholangiopancreatography (MRCP) is crucial for diagnosing hepatobiliary and pancreatic diseases.
• Accelerating MRCP acquisition times is essential to improve patient comfort and reduce motion artifacts.
• Deep learning (DL) shows promise for enhancing medical image reconstruction.

Purpose of the Study:

• To evaluate the efficacy of deep learning-based (DL) reconstruction for accelerating MRCP acquisitions at 3 Tesla (T) and 0.55T.
• To compare DL reconstruction performance against conventional methods like parallel imaging (PI) and compressed sensing (CS).

Main Methods:

• Trained DL reconstruction models using supervised (SV) and self-supervised (SSV) strategies with retrospectively undersampled 3T MRCP data.
• Acquired conventional twofold accelerated MRCP scans from 35 healthy volunteers at 3T and 0.55T.
• Evaluated DL reconstructions using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) metrics, and tested with prospective undersampling.

Main Results:

• DL reconstructions reduced average MRCP acquisition time from 599/542 seconds to 255/180 seconds at 3T/0.55T.
• DL methods achieved higher PSNR and SSIM compared to PI and CS in both retrospective and prospective undersampling.
• Image quality, including sharpness and visibility of hepatobiliary ducts, was preserved with DL reconstructions.

Conclusions:

• Deep learning-based reconstruction effectively accelerates MRCP scans by factors of 2.4 (3T) and 3.0 (0.55T).
• DL methods maintain diagnostic image quality comparable to conventional, longer acquisitions.
• This approach offers a significant advancement for faster and more efficient MRCP imaging across different field strengths.