Related Experiment Video
Updated: Oct 20, 2025
07:59
Training Dogs for Awake, Unrestrained Functional Magnetic Resonance Imaging
Published on: October 13, 2019
7.8K
A deep learning approach for synthetic MRI based on two routine sequences and training with synthetic data
Elisa Moya-Sáez1, Óscar Peña-Nogales2, Rodrigo de Luis-García2
1Laboratorio de Procesado de Imagen, Universidad de Valladolid, Valladolid, Spain. Electronic address: http://www.lpi.tel.uva.es.
Computer Methods and Programs in Biomedicine
|September 15, 2021
Summary
This study introduces a novel learning-based method for generating quantitative magnetic resonance imaging (MRI) parametric maps from standard clinical scans. This approach enables faster, more accessible quantitative MRI in routine clinical practice.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Neuroimaging
Background:
- Quantitative magnetic resonance imaging (MRI) provides valuable diagnostic information but is often time-consuming and requires specialized protocols.
- Existing synthetic MRI methods lack integration into clinical scanners due to specific sequence requirements.
- There is a need for low-cost, clinically compatible methods for quantitative MRI.
Purpose of the Study:
- To develop a learning-based approach for computing T1, T2, and PD parametric maps from routinely acquired T1- and T2-weighted MRI images.
- To enable quantitative MRI using standard clinical imaging protocols and reduce acquisition time.
- To synthesize realistic weighted images from computed parametric maps.
Main Methods:
- A convolutional neural network (CNN) was trained using a synthetic dataset of 120 brain volumes generated from the BrainWeb tool.
- The CNN learns an end-to-end mapping from input T1- and T2-weighted images to T1, T2, and PD parametric maps.
- Conventional weighted images were analytically synthesized from the parametric maps; the network can be fine-tuned with actual clinical data.
Main Results:
- The approach accurately computed parametric maps from synthetic data with normalized squared errors below 1%.
- Realistic parametric maps were generated from actual MRI acquisitions, showing high correlation (>0.95) with literature values and relaxometry-derived maps.
- Synthesized weighted images demonstrated visual realism with mean square errors below 9% and structural similarity indices above 0.90.
Conclusions:
- The proposed CNN approach successfully generates realistic parametric maps and weighted images using only two standard input images acquired in under 8 minutes.
- Synthetic data is crucial for achieving acceptable performance, though fine-tuning with actual clinical data further improves results.
- This method offers a viable solution for quantitative MRI within clinical time constraints and for synthesizing additional weighted images.
