Transfer learning enhanced generative adversarial networks for multi-channel MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning generative adversarial networks (GANs) require extensive data for optimal MR image reconstruction.
• Clinical settings often lack sufficient raw patient data for training these models.
• There is a critical need to improve network generalizability with limited sample sizes.

Purpose of the Study:

• To explore novel applications of parallel imaging combined with GANs (PI-GAN) and transfer learning for MR image reconstruction.
• To evaluate the effectiveness of transfer learning in enhancing model generalizability with small datasets across various clinical scenarios.
• To assess performance improvements in different anatomies and under varying acceleration factors.

Main Methods:

• A PI-GAN model was pre-trained on public brain MRI data.
• The pre-trained model was fine-tuned for specific applications: brain tumors, knee, and liver imaging.
• The model was further tested with different k-space sampling masks and acceleration factors (AFs) of 2 and 6.

Main Results:

• Transfer learning reduced artifacts and improved image smoothness for brain tumor patients.
• For knee and liver imaging, transfer learning outperformed PI-GAN trained on smaller datasets.
• Reconstruction performance improved with transfer learning for both AF=2 and AF=6, with AF=2 yielding better results.
• Transfer learning addressed inconsistencies between training and testing datasets, enhancing generalization.

Conclusions:

• Transfer learning significantly improves the generalizability of deep learning models for MR image reconstruction with limited data.
• This approach is effective across diverse anatomies and sampling strategies, addressing a key challenge in clinical translation.
• Pre-training and fine-tuning offer a viable solution for data-scarce environments in medical imaging AI.