4474
One-Shot Learning for CEST-Centered Multiparametric MRI: Training Neural Network with One Single Scan1Department of Electronic Science, Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, School of Electronic Science and Engineering, National Model Microelectronics College, Xiamen University, Xiamen, China, 2Departments of Radiology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong, China, 3Department of Magnetic Resonance Imaging, The First Affiliated Hospital of Zhengzhou University, Zhengzhou University, Zhengzhou, China, 4Clinical and Technical Support, Philips Healthcare, Beijing, China
Synopsis
Keywords: CEST / APT / NOE, CEST & MT
Motivation: Multiparametric imaging offers comprehensive information. However, its practical application is hindered by extended scanning times.
Goal(s): To develop a CEST-centered multiparametric approach capable of producing multiple quantitative maps.
Approach: ResNet was utilized to simultaneously quantify amide, NOE, MT, DS, B0, T1 and T2. By incorporating a reweighting scheme in conjunction with transfer learning, we demonstrate one single scan is adequate to train a well-performing neural network. The robustness and generalizability of the proposed method were validated using multicenter data.
Results: The proposed method outperformed state-of-the-art CEST deep learning method, providing more accurate quantification results, all while requiring a limited amount of training data.
Impact: The proposed method has the potential to establish a CEST-centered multiparametric approach, eliminating the need for multiple scanning protocols and, consequently, reducing scan time.
Introduction
Multiparametric MRI (mpMRI) offers insights into potential pathologies by combining various tissue contrast parameters. Chemical exchange saturation transfer (CEST) MRI, as a versatile technique, can provide valuable information regarding endogenous metabolites as well as cell microenvironments. Nowadays, deep learning has been introduced to the CEST community to provide quantification with high accuracy, increased speed, and more comprehensive outcomes, such as deepCEST, ANNCEST, and CEST-MRF.1,2,3,4,5 However, the neural network adopted in these methods is the classical feedforward neural network, which is prone to overfitting issues and can easily become trapped in local optima during the training process. Furthermore, the training data for these methods is generated through numerical simulations, which may deviate from real-world applications due to simplified simulation models.In this study, Reweighted ResNet CEST (RRC) was developed to provide multiparametric quantification, including amide, Nuclear Overhauser Effect (NOE), Magnetization Transfer (MT), Direct Saturation (DS), as well as B0, T1, and T2. By incorporating a reweighting scheme in combination with transfer learning, we demonstrate that training data extracted from one single scan is sufficient to train a high-performing neural network. The results of the proposed method were compared with a state-of-the-art method and validated across multiple centers.
Methods
The neural network adopted in this study consists of a 1D convolutional network with residual structure, of which the input is the Z-spectrum, and the output includes multiparametric quantification, including amide, NOE, MT, DS, B0, T1, and T2. The flowchart of the proposed method is illustrated in Figure 1. The training data were acquired from brain tumor patients using a multi-slice CEST protocol at 3T, with 2μT and 2s saturation pulse, covering a frequency range from -5ppm to 5ppm.To address data imbalance in the real acquired datasets, we developed a reweight-balancing scheme that includes Gaussian distribution fitting and loss reweight. Transfer learning was incorporated to diminish the need for a large training dataset. Multicenter data were acquired to validate the robustness and generalizability of the proposed method.Results and Discussion
Figure 2 presents the training curves of RRC and ANN, as well as their performance as a function of the layer number. The results reveal that ResNet provides superior quantification accuracy compared to ANNCEST. This improvement can be attributed to the advantages of the residual structure in mitigating overfitting issues. The data imbalance problem is effectively resolved by the reweight-balancing scheme, resulting in enhanced quantification outcomes for RRC.Figure 3 displays the quantification results and index metrics for ANNCEST and RRC. The results indicate that the proposed method can offer results with higher fidelity, particularly in the tumor regions. The performance of ANNCEST and RRC on 4 patients and validation at different centers is summarized in Table 1. The results show that our proposed method outperforms the existing method. The time consumption of ANNCEST and RRC is less than 40 seconds, which is much faster compared to fitting methods such as Lorentzian fitting and PLOF, which take minutes.
While deep learning has been introduced to CEST quantification, the training data used in previous studies were generated using numerical simulations, which may compromise quantification accuracy and generalizability due to the simplified models. Here, we demonstrate the feasibility of training a neural network using only one single scan of real acquired data using RRC. Moreover, few studies have attempted to quantify T1 and T2 relaxation times from Z-spectrum. Instead, previous studies required additional T1 and T2 information for accurate quantification.1,4 Our proposed method demonstrates that T1 and T2, as well as various CEST parameters and B0, can be accurately quantified from the Z-spectrum.
Conclusion
In this study, we developed a CEST-centered multiparametric strategy named reweighted ResNetCEST, enabling the rapid quantification of multiple parameters with minimal training data requirements. This approach has the potential to eliminate the need for multiple scanning protocols, ultimately reducing scan time.Acknowledgements
This work is supported by the National Natural Science Foundation of China, Grant/Award Number:82302151; Shenzhen Science and Technology Program, Grant/Award Number: JCYJ20220818101213029; Fujian Province Science and Technology Project, Grant/Award Number: 2022J05013; Xiamen University Nanqiang Outstanding Talents Program.References
1.Huang J, Lai JHC, Tse KH, Cheng GWY, Liu Y, Chen Z, Han X, Chen L, Xu J, Chan KWY. Deep neural network based CEST and AREX processing: Application in imaging a model of Alzheimer's disease at 3 T. Magnetic Resonance in Medicine 87, 1529-1545 (2022).
2.Chen L, Schar M, Chan KWY, Huang J, Wei Z, Lu H, Qin Q, Weiss RG, van Zijl PCM, Xu J. In vivo imaging of phosphocreatine with artificial neural networks. Nature Communications 11, 1072 (2020).
3.Kim B, Schar M, Park H, Heo HY. A deep learning approach for magnetization transfer contrast MR fingerprinting and chemical exchange saturation transfer imaging. Neuroimage 221, 117165 (2020).
4.Perlman O, Ito H, Herz K, Shono N, Nakashima H, Zaiss M, Chiocca EA, Cohen O, Rosen MS, Farrar CT. Quantitative imaging of apoptosis following oncolytic virotherapy by magnetic resonance fingerprinting aided by deep learning. Nat Biomed Eng 6, 648-657 (2022).
5.Cohen O, Yu VY, Tringale KR, Young RJ, Perlman O, Farrar CT, Otazo R. CEST MR fingerprinting (CEST-MRF) for brain tumor quantification using EPI readout and deep learning reconstruction. Magnetic Resonance in Medicine 89, 233-249 (2023).
Figures
