Purpose: Quantitative T2* mapping is important for the diagnosis and management of knee diseases. However, derivation of T2* maps typically requires acquisition of multi-echo images, which is time consuming. In this study, we present a new paradigm to predict T2* map from two conventional images obtainable in a standard clinical MRI examination, thus eliminating additional data acquisition for quantitative MRI.
Methods: A deep learning framework is proposed to predict T2* map from a T1-weighted image and a T2*-weighted image. The principle is validated on knee MRI. 1344 T2* maps from 56 subjects are retrospectively collected. Every T2* map was extracted from six T2*-weighted images acquired with different echo time (of 0.032, 4.4, 8.8, 13.2, 17.6, and 22ms); in addition, a T_1-weighted image was acquired. To derive T2* map from two input images, a self-attention convolutional neural network is employed. Here, a hierarchical network architecture is adopted; global and local shortcuts are equipped to facilitate residual learning; and attention mechanism is integrated to make efficient use of non-local information. Separate deep neural networks are trained, each accepting T2*-weighted image obtained with a specific TE. In training, a mixed function of l1 loss and SSIM loss is employed, and network parameters are updated using the Adam algorithm. All 1344 images are utilized for training and testing in a six-fold cross-validation strategy. The prediction result is quantitatively evaluated within the region of interest (i.e. cartilage manually segmented).
Results: Using well-trained models, T2* maps are predicted from single T2*-weighted images (acquired using TE of 4.4, 8.8, 13.2, or 17.6ms) along with corresponding T1-weighted image. High accuracy has been achieved. The averaged L1 error with ROI is 0.13.
Conclusion: A data-driven strategy is proposed for T2* mapping from two conventional MR images, which has a potential to eliminate additional data acquisition for quantitative MRI.