Abstract
Quantitative T2 mapping provides absolute physical values for tissue characterization, yet its clinical adoption is hindered by long scan and reconstruction times. This study introduces an end-to-end deep learning framework using a Residual Network (ResNet) to accelerate T2 mapping from single-shot OverLapping-Echo Detachment (OLED) MRI. The proposed method bypasses computationally expensive iterative optimization, reducing reconstruction times from several minutes to tens of milliseconds. Results on simulated and in vivo human brain data demonstrate that the ResNet approach achieves higher accuracy, better noise suppression, and fewer artifacts than traditional methods, offering a path toward real-time quantitative MRI.
Introduction
Standard MRI typically produces qualitative "weighted" images that vary across scanners. In contrast, quantitative T2 mapping provides absolute physical values, making it a powerful tool for diagnosing neurological conditions like stroke and epilepsy. However, the slow nature of T2 mapping remains a bottleneck. While the OverLapping-Echo Detachment (OLED) sequence was developed to acquire data in a single shot, the resulting signals are mathematically complex due to overlapping echoes. Traditional reconstruction relies on non-linear optimization algorithms that are too slow for real-time clinical workflows.
Methods
The researchers implemented an 8-layer Deep Residual Network (ResNet) to perform end-to-end reconstruction. Unlike standard CNNs, ResNets utilize "shortcut connections" to learn residual functions, which helps mitigate the vanishing gradient problem and improves training efficiency for image regression tasks.
The network takes a two-channel input (real and imaginary parts of the complex OLED signal) and directly outputs a single-channel T2 map. Training was conducted using 100 pairs of simulated OLED images and T2 maps generated from a BrainWeb phantom. To ensure robustness, the model was trained with varying sequence parameters and added Gaussian noise. In vivo validation was performed on healthy volunteers using a 3T scanner, with a multi-scan Spin-Echo (SE) sequence serving as the gold standard.
Results
The ResNet approach achieved a massive speedup, slashing reconstruction times from minutes to tens of milliseconds. Beyond speed, the deep learning model demonstrated superior image quality. In regions with long T2 values, such as cerebrospinal fluid (CSF), the ResNet provided higher accuracy and better resolution than iterative methods. Quantitative Region of Interest (ROI) analysis showed that ResNet's T2 values closely matched the gold standard, with a maximum deviation of only 7.0%, compared to 10.7% for the traditional echo-detachment method. Furthermore, the ResNet effectively suppressed "lattice-shaped" artifacts caused by signal modulation streaks.
Discussion and Conclusion
A key insight of this work is that a network trained entirely on simulated data can generalize effectively to real human brain data, eliminating the need for large clinical training sets. While the current OLED sequence is limited by a 128×128 matrix and some T2 decay blurring, the ResNet framework significantly improves the feasibility of real-time quantitative imaging. Future work will focus on higher-resolution acquisition schemes and extending this approach to other complex MRI sequences where traditional reconstruction remains a bottleneck.
Summary updated