Accurate T2 mapping using multi-echo spin-echo data is a time-consuming process due to stimulated echo correction. In this study, we developed an artificial neural network for real-time T2 mapping. The training dataset using both in-vivo data and model-based synthetic data demonstrated the best performance. The resulting T2 map shows mean T2 errors of less than 0.3 ms with minimal computation time (less than 1 sec as opposed to 8.3 hours for conventional method). An additional algorithm was developed to ensure the fidelity of the T2 map at the cost of slightly increased computation time.
[Artificial neural network]
As depicted in Figure 2, the T2 mapping ANN was designed to have 5 different fully connected layers (FCNs) with each FCN except the output layer connected to a rectified linear unit (ReLU). The input layer receives multi-echo spin-echo signals (six echoes in this work) normalized by the first echo magnitude. The outputs of the ANN are E2 and B1, where E2 = exp(-TE1/T2). The T2 value is calculated by substituting TE1 into E2, where TE1 is the first echo-time. The training cost $$$L(\theta)$$$ was defined as $$L(\theta)=\frac{1}{N}∑_{\:i=1}^{\:N}[(E_{2,label}\:(i)-E_{2,output}\:(i;\theta))^{2}+(B_{1,label}\:(i)-B_{1,output}\:(i;\theta))^{2}]$$, where N is the size of data and $$$\theta$$$ are the trainable parameters of the ANN. This approach of using E2 instead of T2 converts the input range from [0,∞) in T2 to (0,1] in E2 and, therefore, enables the ANN to be applied for any TE. Additionally, this approach overcomes the limited T2 range of in-vivo data because a wide range of E2 can be acquired by changing TE. For quality assurance of the T2 mapping (QA process), voxels with large residual error (RMSE>0.02) were reprocessed using EPGSLR dictionary-matching with the estimated T2 from the ANN as an initial point to facilitate an efficient search.
[In-vivo and synthetic data]
For a training set of the ANN, in-vivo data and the EPGSLR fitted E2 and B1 results were used. 2D multi-echo spin-echo data from 5 subjects (1 for test set) were acquired at 3T using the following parameters: four different echo spacings = 9.5, 20, 32 and 46 ms, echo train length = 6, slice thickness = 2 mm, resolution = 1.5×1.5 mm2 and number of slices = 50. Total voxel numbers for training were 120,000 per subject. Synthetic data were generated for different E2 and B1 values using the EPGSLR model. The values were uniformly distributed for E2 ([0,1]) and B1 ([0.5,1.5]).
[Evaluation]
To test the dependencies on the training data composition and size, the ANN was trained with various training sets (in-vivo-only, synthetic-only and both) and a T2 error was evaluated by the difference from the actual T2 values determined by the EPGSLR curve-fitting. For the best performing ANN, the T2 map was compared with that from EPGSLR curve-fitting and EPGSLR dictionary-matching (T2 resolution = 0.5 ms). The processing time was calculated under the same computing environment (MATLAB; 8 core 2.4 GHz CPU). Additionally, EPGSLR signals (SNR(TE1) = 200, repetition = 1000) were generated to compare T2 errors for broad ranges of T2 (10-300 ms) and B1 (0.5-1.5).
1. Jackson GD, Connelly A, Duncan JS, Grünewald RA, Gadian DG. Detection of hippocampal pathology in intractable partial epilepsy: Increased sensitivity with quantitative magnetic resonance T2 relaxometry. Neurology 1993;43(9):1793-1799.
2. Pell GS, Briellmann RS, Waites AB, Abbott DF, Lewis DP, Jackson GD. Optimized clinical T2 relaxometry with a standard CPMG sequence. J Magn Reson Imaging 2006;23(2):248-252.
3. Dunn TC, Lu Y, Jin H, Ries MD, Majumdar S. T2 relaxation time of cartilage at MR imaging: Comparison with severity of knee osteoarthritis. Radiology 2004;232(2):592-598.
4. Hennig J. Multiecho imaging sequences with low refocusing flip angles. Journal of Magnetic Resonance (1969) 1988;78(3):397-407.
5. Lebel RM, Wilman AH. Transverse relaxometry with stimulated echo compensation. Magn Reson Med 2010;64(4):1005-1014.
6. McPhee KC, Wilman AH. Transverse relaxation and flip angle mapping: Evaluation of simultaneous and independent methods using multiple spin echoes. Magn Reson Med 2017;77(5):2057-2065.
7. Bishop CM, Roach CM. Fast curve fitting using neural networks. Review of Scientific Instruments 1992;63(10):4450-4456.
8. Hornik K. Approximation capabilities of multilayer feedforward networks. Neural Networks 1991;4(2):251-257.