Transfer learning in deep neural network-based receiver coil sensitivity map estimation | JoVE Visualize

Transfer learning in deep neural network-based receiver coil sensitivity map estimation

Madiha Arshad¹, Mahmood Qureshi², Omair Inam²

¹Medical Image Processing Research Group (MIPRG), Department of Electrical and Computer Engineering, COMSATS University, Islamabad, Pakistan. engr_madiha88@yahoo.com.

Magma (New York, N.Y.)|March 27, 2021

View abstract on PubMed

Summary

Transfer learning enhances deep learning for Magnetic Resonance Imaging (MRI) coil sensitivity maps. This approach overcomes data scarcity and improves SENSitivity Encoding (SENSE) reconstruction accuracy, reducing retraining costs.

Area of Science:

Medical Imaging
Magnetic Resonance Imaging

Background:

Accurate receiver coil sensitivity maps are crucial for parallel MRI algorithms like SENSitivity Encoding (SENSE).
Deep learning methods for sensitivity map estimation face challenges with training dataset size and generalization.
Retraining deep learning models from scratch due to data mismatches is inefficient and costly.

Purpose of the Study:

To address data scarcity and generalization issues in deep learning-based receiver coil sensitivity map estimation.
To evaluate a transfer learning approach for estimating 3T receiver coil sensitivity maps using a model pre-trained on 1.5T data.

Main Methods:

A transfer learning strategy involving end-to-end fine-tuning of a pre-trained U-Net model was employed.
The U-Net model, initially trained on 1.5T sensitivity maps, was assessed for its generalization capabilities at 3T.
The pre-trained U-Net was subsequently fine-tuned to estimate 3T receiver coil sensitivity maps.

Main Results:

The fine-tuned U-Net model successfully estimated 3T receiver coil sensitivity maps.
Quantitative metrics including Peak Signal-to-Noise Ratio and Root Mean Square Error were evaluated.
Analysis of central line profiles in SENSE reconstructed images demonstrated the effectiveness of the estimated maps.

Conclusions:

The proposed transfer learning method effectively estimates receiver coil sensitivity maps for 3T MRI.
This approach enables successful SENSE reconstruction, overcoming limitations of data scarcity and generalization in deep learning models.
End-to-end fine-tuning provides a cost-effective and time-efficient solution for improving MRI reconstruction accuracy.