A deep learning approach for fast muscle water T2 mapping with subject specific fat T2 calibration from multi-spin-echo acquisitions

Abstract

This work presents a deep learning approach for rapid and accurate muscle water T₂ with subject-specific fat T₂ calibration using multi-spin-echo acquisitions. This method addresses the computational limitations of conventional bi-component Extended Phase Graph fitting methods (nonlinear-least-squares and dictionary-based) by leveraging fully connected neural networks for fast processing with minimal computational resources. We validated the approach through in vivo experiments using two different MRI vendors. The results showed strong agreement of our deep learning approach with reference methods, summarized by Lin's concordance correlation coefficients ranging from 0.89 to 0.97. Further, the deep learning method achieved a significant computational time improvement, processing data 116 and 33 times faster than the nonlinear least squares and dictionary methods, respectively. In conclusion, the proposed approach demonstrated significant time and resource efficiency improvements over conventional methods while maintaining similar accuracy. This methodology makes the processing of water T₂ data faster and easier for the user and will facilitate the utilization of the use of a quantitative water T₂ map of muscle in clinical and research studies.

Figure 1

Schematization of the NN models and the training pipeline used for training. (A) Fat-Net and (B) Muscle-Net have the same architecture: 4 fully connected hidden layers with a progressively decreasing number of neurons. (C) During training, simulated data are augmented by injecting white gaussian noise with different noise variances. Signals are normalized before being fed to the NN.

Figure 2

Processing pipeline used to analyze the data. Image pre-processing (top panel) and data analysis (bottom panel) steps used to produce the parametric maps for each subject: T_2w, FF and B₁⁺ with subject specific T_2f. calibration.

Figure 3

Example of T_2w, FF and B₁⁺ quantitative maps obtained in the thigh from EPG fitting of MESE data acquired with Siemens scanner. The proposed deep learning approach (central column, red frame) is presented along with the maps obtained using the dictionary (left column) and NLSQ (right column) reference fitting methods.

Figure 4

Example of T_2w, FF and B₁⁺ quantitative maps obtained in the thigh from EPG fitting of MESE data acquired with Philips scanners. The proposed deep learning approach (central column, red frame) is presented along with the maps obtained using the dictionary (left column) and NLSQ (right column) reference fitting methods.

Figure 5

Bland–Altman plots reporting the comparison of T_2w values obtained using the prosed DL approach with the reference dictionary (left) and NLSQ (right) EPG fitting methods for data acquired using Simens (top panel) and Philips (bottom panel) scanners. The width of the LOA is reported for each BA plot along with the Lin’s concordance coefficient. The different colors represent muscle T_2w values belonging to different subjects.

Figure 6

Scatter plots of estimated T_2w values against Dixon fat fraction values. The plots are reported separately for Siemens (top row) and Philips data (bottom row). A plot for each EPG fitting procedure investigated is reported: the proposed DL approach and the reference dictionary and NLSQ methods. The different colors represent values belonging to different subjects.