Linear fine-tuning: a linear transformation based transfer strategy for deep MRI reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Transfer Learning

Background:

• Deep learning-based magnetic resonance imaging (MRI) reconstruction commonly uses fine-tuning (FT), initializing models with pre-trained weights and updating them with target domain data.
• Direct full-weight updates in FT risk catastrophic forgetting and overfitting, limiting its effectiveness in MRI reconstruction.

Purpose of the Study:

• To develop a zero-weight update transfer strategy for MRI reconstruction.
• To preserve pre-trained knowledge and reduce overfitting during transfer learning.

Main Methods:

• Propose Linear Fine-Tuning (LFT), a novel transfer strategy assuming a linear transformation between source and target domain weights.
• LFT introduces learnable scaling and shifting (SS) factors, keeping pre-trained weights fixed and only updating SS factors during transfer.

Main Results:

• LFT demonstrated superior performance over FT in transfer scenarios involving different contrasts, slice directions, and anatomical structures.
• LFT significantly reduced artifacts and improved peak signal-to-noise ratio (up to 2.06 dB) in reconstructed MRI images.
• Performance gains were most pronounced when the target domain had limited training data.

Conclusions:

• LFT effectively addresses catastrophic forgetting and overfitting in MRI reconstruction transfer learning.
• The strategy reduces data dependency in the target domain, shortening model development cycles.
• LFT enhances the clinical applicability of deep learning models for MRI reconstruction.