Only-Train-Once MR Fingerprinting for Magnetization Transfer Contrast Quantification

Beomgu Kang, Hye-Young Heo, HyunWook ParkMICCAI 2022 · OpenReview

Abstract

Magnetization transfer contrast magnetic resonance fingerprinting (MTC-MRF) is a novel quantitative imaging technique that simultaneously measures several tissue parameters of semisolid macromolecule and free bulk water. In this study, we propose an Only-Train-Once MR fingerprinting (OTOM) framework that estimates the free bulk water and MTC tissue parameters from MR fingerprints regardless of MRF schedule, thereby avoiding time-consuming process such as generation of training dataset and network training according to each MRF schedule. A recurrent neural network is designed to cope with two types of variants of MRF schedules: 1) various lengths and 2) various patterns. Experiments on digital phantoms and in vivo data demonstrate that our approach can achieve accurate quantification for the water and MTC parameters with multiple MRF schedules. Moreover, the proposed method is in excellent agreement with the conventional deep learning and fitting methods. The flexible OTOM framework could be an efficient tissue quantification tool for various MRF protocols.

Introduction

Magnetization Transfer Contrast (MTC) imaging is a powerful tool for characterizing semisolid macromolecules, such as proteins and lipids, which are critical biomarkers in neurodegenerative diseases like Multiple Sclerosis. While Magnetic Resonance Fingerprinting (MRF) has enabled quantitative MTC imaging, existing deep learning approaches are often constrained by a "one-schedule, one-network" paradigm. Standard fully connected neural networks (FCNNs) are trained for a specific pulse sequence (schedule). Any modification to the scan duration or pulse pattern necessitates the generation of new training data and a time-consuming retraining process, limiting clinical flexibility and slowing the adoption of MRF in diverse research settings.

Methods

To overcome this rigidity, the authors propose the Only-Train-Once MR fingerprinting (OTOM) framework. This approach shifts from static mapping to sequence modeling using a Bidirectional Long Short-Term Memory (bi-LSTM) network.

The OTOM framework architecture. (A) The training loop integrates Bloch-McConnell equations with a bi-LSTM quantification network. (B) Pulse sequence diagram showing RF saturation and relaxation delays. (C) Input structure where scan parameters ($p_{scan}$) and signal ($S_{MTC-MRF}$) are fed into the network at each time point.

The core innovation lies in feeding the scan parameters (flip angles, timings, etc.) into the network alongside the MR signal at every time step. This "physics-aware" design allows the bi-LSTM to learn the underlying signal evolution dynamics rather than just memorizing specific signal shapes. Because LSTMs are inherently designed for sequences, the model can generalize to various scan lengths and patterns without retraining.

Results

The OTOM framework was evaluated against conventional schedule-specific networks and traditional dictionary fitting methods using digital phantoms and in vivo data from six healthy volunteers.

Comparison between conventional multi-network approaches and the OTOM framework. (A) Traditional methods require separate networks for different schedules (MTC-MRF1 vs. MTC-MRF2). (B) The proposed OTOM framework uses a single network to accurately quantify tissue parameters across different schedules.

Key findings include:

  • Efficiency: OTOM performed quantification in approximately 110 seconds, a massive improvement over the one-hour requirement for traditional Bloch fitting methods.
  • Accuracy: The Mean Absolute Error (MAE) for tissue parameters (such as exchange rate $k_{mw}$ and macromolecular fraction $M_0^m$) in digital phantoms was nearly identical to FCNNs trained specifically for those schedules.
  • Robustness: In vivo results showed high correlation coefficients (0.89–0.99) with established methods, demonstrating the model's reliability in real human biology.
  • Flexibility: The single OTOM network successfully processed multiple "pseudo-random" schedules with varying lengths (e.g., 30 vs. 40 scans) and patterns that it had never seen in those exact configurations during training.

Conclusion

The OTOM framework provides a flexible and efficient solution for quantitative MTC-MRF. By leveraging the temporal processing capabilities of RNNs and integrating scan parameters into the learning process, it eliminates the need for repeated network training. This approach could serve as a universal decoder for various MRF protocols, significantly lowering the barrier to clinical adoption. Future work may extend this framework to more complex multi-pool tissue models and investigate its performance under diverse noise and artifact conditions.

Summary updated

Abstract

Magnetization transfer contrast magnetic resonance fingerprinting (MTC-MRF) is a novel quantitative imaging technique that simultaneously measures several tissue parameters of semisolid macromolecule and free bulk water. In this study, we propose an Only-Train-Once MR fingerprinting (OTOM) framework that estimates the free bulk water and MTC tissue parameters from MR fingerprints regardless of MRF schedule, thereby avoiding time-consuming process such as generation of training dataset and network training according to each MRF schedule. A recurrent neural network is designed to cope with two types of variants of MRF schedules: 1) various lengths and 2) various patterns. Experiments on digital phantoms and in vivo data demonstrate that our approach can achieve accurate quantification for the water and MTC parameters with multiple MRF schedules. Moreover, the proposed method is in excellent agreement with the conventional deep learning and fitting methods. The flexible OTOM framework could be an efficient tissue quantification tool for various MRF protocols.
BibTeX
@inproceedings{kang2022only,
  author = {Beomgu Kang and Hye-Young Heo and HyunWook Park},
  title = {{Only-Train-Once MR Fingerprinting for Magnetization Transfer Contrast Quantification}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  url = {https://openreview.net/forum?id=MDnSwe7snj},
}
Figure