Inter-scanner super-resolution of 3D cine MRI using a transfer-learning network for MRgRT

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy

Background:

• Deep-learning super-resolution (SR) enhances 3D MRI spatial resolution for MR-guided radiotherapy (MRgRT).
• Inter-scanner and patient variations degrade SR quality in real-time 3D low-resolution (LR) cine MRI.

Purpose of the Study:

• To present a personalized super-resolution (psSR) network using transfer-learning to address inter-scanner SR challenges in 3D cine MRI.
• To evaluate the impact of psSR on image quality and clinical feasibility for MRgRT applications.

Main Methods:

• Developed a two-stage network: cohort-specific SR (csSR) trained on 53 patients at 1.5 T, followed by psSR using transfer-learning on 5 healthy volunteers at 0.55 T.
• Evaluated image quality using peak-signal-noise-ratio (PSNR) and structure-similarity-index-measure (SSIM).
• Assessed clinical feasibility via liver auto-segmentation using the dice-similarity-coefficient (DSC).

Main Results:

• psSR significantly increased mean PSNR by 57.2% and SSIM by 94.7% compared to cine MRI.
• Liver contouring DSC improved by 15% using psSR MRI.
• Transfer-learning averaged 90 seconds, psSR reconstruction took 4.51 ms/volume, and auto-segmentation was 210 ms.

Conclusions:

• The psSR network substantially enhances 3D cine MRI quality and segmentation accuracy within minutes of transfer-learning.
• This approach effectively overcomes deep-learning's cohort- and scanner-dependency for MRgRT, improving real-time image guidance.