Transfer-learning is a key ingredient to fast deep learning-based 4D liver MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Time-resolved volumetric 4D MRI is crucial for image-guided interventions, but current methods struggle with organ motion due to limitations in resolution and speed.
• Deep learning (DL) offers potential solutions but faces challenges with domain shift, impacting real-world applicability.

Purpose of the Study:

• To evaluate transfer learning (TL) combined with ensembling as a strategy to overcome domain shift in DL-based 4D MRI reconstruction.
• To improve the temporal and spatial resolution and reduce reconstruction time for 4D MRI in interventional settings.

Main Methods:

• Four DL approaches were compared: source-pretrained, target-trained from scratch, fine-tuned, and an ensemble of fine-tuned models.
• A dataset was split into 16 source and 4 target domain subjects to assess performance.
• The study focused on 4D organ motion models, particularly for the liver.

Main Results:

• The ensemble of fine-tuned models (N=10) showed significant improvements (P < 0.001) compared to directly learned models.
• Root mean squared error (RMSE) improved by up to 12%, and mean displacement (MDISP) by up to 17.5%.
• The benefits of TL + Ens were more pronounced with smaller target domain datasets.

Conclusions:

• Transfer learning combined with ensembling effectively mitigates domain shift in 4D MRI reconstruction.
• This strategy significantly reduces prior acquisition time and enhances reconstruction quality.
• TL + Ens is a key advancement for clinical feasibility of 4D MRI in tracking organ motion.