Introduction

DTI is widely used to examine the human brain white matter structures, including their microarchitecture integrity and spatial fiber tract trajectories. Although theoretically, only six diffusion directions are necessary for estimation of the diffusion tensor, much more DWIs from different directions are needed in practice due to the low SNR, resulting in a prolonged acquisition. Such prolonged scans can increase motion artifacts and discomfort of patients. Several techniques have been developed for DTI acceleration, such as parallel imaging based simultaneous multislice (SMS) acquisition [1] and compressed sensing [2-7]. However, the acceleration factor is limited to 2-3 with extensive computation power. The q-DL [8] and other subsequent studies have demonstrated the potential of using machine learning to reduce the q-space data necessary for diffusion image [9-14]. Besides multi-layer perceptrons [8, 9, 13], some used convolutional neural networks [10-12, 14-17]. Using deep learning, we demonstrate for the first time the feasibility to generate FA/MD and fiber tractography using as few as six undersampled DWIs indistinguishable from those obtained using 90 DWIs. SuperDTI is also compared against the state-of-the-art methods for diffusion map reconstruction and fiber tracking.

Methods

In the deep learning network, the goal is to learn the nonlinear relationship $$$F$$$ between input $$$x$$$ and output $$$y$$$, which is represented as $$$y=F(x;Θ)$$$, where $$$Θ$$$ is the DL parameters to be learned during training. Learning of such a mapping is achieved through minimizing a loss function between the network prediction and the corresponding ground truth data based on some training data. Our goal is to obtain the mapping between DWIs (input) and the desired maps (output) using deep learning while by-passing the conventional tensor model. During training, a reduced number of DWIs $$$x_i$$$ are used as the inputs and the corresponding true maps $$$y_i$$$ (FA, MD or Eigenvectors obtained by fitting all 90 b = 1000 s/mm² images to the tensor model) as the output. We learn deep learning network parameters $$$Θ$$$ that minimize the loss function: $$L(Θ)=\frac{1}{n} \sum_{i=1}^n ‖F(x_i;Θ)-y_i ‖^2 (1),$$ which is the mean-square error (MSE) between the network output and the ground truth maps (n is the number of training dataset). The proposed network comprises several layers of a skip-connection-based convolution-deconvolution network, which learns the residual between its input and output, as shown in Figure 1. In each layer, n64k3s1p1 (p’1) indicates 64 filters of kernel size 3 × 3 with a stride of 1 and padding of 1 (a truncation of 1). Except for the last layer, each (de)convolutional layer is followed by a ReLU unit. In testing, acquired DWIs are fed into the network with learned $$$Θ$$$ to generate the desired maps $$$F(x_t;Θ)$$$. The DTI model is not needed during testing, thus avoiding the model fitting error. Data from 40 subjects were used for training and validation, and the data from the other 10 subjects were used for testing. The subjects were randomly selected from the Human Connectome Project [18]. Each dataset includes 18 non-DWIs and 270 DWIs in three different b values: 1000, 2000, and 3000 s/mm2 and 90 diffusion directions. The reduced DWIs were uniformly selected with b=1000 s/mm². The hardware specification is CPU i9 7980XE (18 cores 36 threads @4.2GHz); GPU 2x NVIDIA Quadro P6000 (24GB each); Memory 128 GB. The training takes around 10 hours. In contrast, it takes only 0.005 seconds to generate each desired map using the learned network.

Results

Figure 1 shows the framework of SuperDTI. Figures 2 shows representative FA maps from 6 DWIs generated using the conventional tensor model fitting (MF), multi-layer perceptron (MLP) (similar to [8]), block-matching and 4D filtering (BM4D) denoising algorithm [18, 19], the proposed deep learning method (SuperDTI), and the proposed method with 2 × k-space undersampling (SuperDTI+). Figure 3 shows representative MD maps from 3 DWIs using different methods. While other results became noisy or blurry in such an extreme case, the FA maps generated by our proposed method showed no apparent degradation. The generated FA/MD maps were quantitatively evaluated using the PSNRs, SSIMs, and NMSEs, which show good performance for SuperDTI even with only 6 DWIs for FA and 3 DWIs for MD maps. It is seen that the difference between the MD maps is far less apparent than that between the FA maps because the MD calculation is less sensitive to noise. In Figure 4, fiber tracking results using the proposed method better preserve the morphology of three major white matter tracts in the brain than the others in the extreme 6 DWIs situation.

Conclusion

In this abstract, we present SuperDTI for superfast diffusion tensor imaging and fiber tractography using deep learning with as few as six undersampled DWIs (up to 30-fold). Such a significant reduction in scan time will allow the inclusion of DTI into clinical routine for many potential applications. Future studies will use patient data for evaluating quantification accuracy and diagnostic performance.

Acknowledgements

This work is supported in part by the National Institute of Health Brain Initiative R01EB025133.

References

[1] Barth, Breuer et al. MRM. 2016 [2] Landman, Wan et al. SPIE. 2010 [3] Menzel, Tan et al. MRM. 2011 [4] Michailovich, Rathi et al. TMI. 2011 [5] Landman, Bogovic et al. NeuroImage. 2012 [6] Wu, Zhu et al. MRM. 2014 [7] Shi, Ma et al. MRM. 2015 [8] Golkov, Dosovitskiy et al. TMI. 2016 [9] Poulin, Cote et al. MICCAI. 2017 [10] Gong, He et al. ISMRM. 2018 [11] Li, Zhang et al. ISMRM. 2018 [12] Aliotta, Nourzadeh et al. ISMRM. 2019 [13] Gibbons, Hodgson et al. Medical Physics. 2019 [14] Li, Zhang et al. MRM. 2019 [15] Li, Gong et al. IEEE Access. 2019 [16] Lin, Gong et al. Medical Physics. 2019 [17] Tian, Bilgic et al. NeuroImage. 2020 [18] Van Essen, Ugurbil et al. NeuroImage. 2012 [19] Dabov, Foi et al. IEEE Trans. Image Process. 2007 [20] Maggioni, Katkovnik et al. IEEE Trans. Image Process. 2012