Introduction

With great potential in probing tissue cellularity and detecting of acute brain stroke non-invasively, DWI has been widely used for clinical assessment. Compared with single-shot DWI, we can be benefited from multi-shot DWI in aspects like less distortion, sub-millimeter resolution and shorter echo times, while hampered by phase errors and longer scanning and reconstruction time. Hence, we proposed a novel method to jointly recover high resolution multiple b-value DWIs with single-shot data, showing high-quality Apparent Diffusion Coefficients (ADC) mapping for clinical generalization.

Methods

Datasets
The study recruited a total of 32 healthy volunteers, each contains 16 slices. Subjects were randomly grouped into training (subjects/slices = 22/352), validation (5/80) and testing sets (5/80) respectively. Tumor patients with lymphatic metastasis were recruited additionally for testing. All data were acquired on a 3.0 T MRI scanner (Ingenia CX, Philips Healthcare, Best, the Netherlands) equipped with a 32-channel head coil. MS-DWI was obtained using four-shot interleaved EPI sequence without navigator, with parameters as follows: TE, 75 ms; TR, 2800 ms; matrix size, 228x228; slice thickness, 4 mm; partial Fourier factor, 0.702; voxel size, 1.0x1.0 mm². SS-DWI and MS-DWI sequences, including four b-values of 0, 500, 750 and 1000 s/mm², were applied in the anterior-posterior direction. SS-DWI was performed with the same EPI sequence as MS-DWI, but using SENSE with an acceleration factor of 2.
MUSE Reconstruction
We used MUSE¹ to generate ground truth images in our network by estimating the motion-induced phase variations among multiple shots, and jointly calculating the magnitude signals of aliased voxels from all segments.
Network Setup
We extended a cascaded CNN (Figure 1) into a multi-channel setup: $$\sum_1^mu_{j}=argmin_{u_{1,2,3,4}}\sum_1^m||Eu_j-f_j||_2^{2}+\lambda_j||u_j-D(u_j;\theta_j)||_2^2$$where $$$D(u_j;\theta_j)$$$is a residual network for the j^th channel with the parameter $$$\theta_j$$$, $$$u_j$$$ represents the predicted image of j^th b-value and $$$f_j$$$ is the undersampled k-space data. The proposed model is an end-to-end multi-channel network with a data consistency layer and a de-aliasing layer^2,3. Single-shot images with b-values of 0, 500, 750 and 1000 s/mm² and corresponding coil sensitivity maps are fed into the network. Subsequently, real and imaginary parts of images are reconstructed through two separate networks. The final outputs are joined to form a coil-combined complex image. Each de-aliasing layer contains five convolutional layers, including 64 filters with a size of 3x3 and a stride of 1. The first four convolutional layers are followed by batch normalization and ReLU. The last layer is without ReLU. The total losses of four channels are combined together and back propagated to get the best weightings between data fidelity and de-aliasing of each channel. The model was implemented with Tensorflow using 4 GPUs of NVIDIA Tesla DGX (4 cores, each with 32 GB memory). The filter weights of each layer were initialized by Xavier⁴ and updated using ADAM optimizer algorithm⁵ with a fixed learning rate of 10^-3. The training epoch was set to 1000, and lasting about 50 hours.
Evaluation Metrics
The proposed network was compared to ESPIRiT⁶ and ResNet. The reconstruction performance was measured by Peak SNR (PSNR) and the Structure SIMilarity index (SSIM). ADC maps were obtained by fitting a pixel-wise mono-exponential function using DWI with different b-values.

Results

Compared with SS-DWI, the proposed method achieved much higher spatial resolution with single-shot scanning time. (Figure 2 and Figure 4) The aliasing artifact was quite clear in the error map of ESPIRiT reconstruction (SSIM = 0.58 for b = 1000 s/mm²). Single-channel ResNet converged faster during training, while showing severe over-smoothing effects with lower SSIM and PSNR for each b-value image compared with the proposed method (SSIM = 0.62 for b = 1000 s/mm²). By removing aliasing artifacts, the proposed method achieved accurate reconstruction, outperforming ESPIRiT and single-channel ResNet among all quality metrics for all b-values (SSIM = 0.75 for b = 1000 s/mm²). Tumor patients with lymphatic metastasis were further evaluated for model generalization. (Figure 3, Figure 4 and Figure 5) It should be noted that no pathological images were used for training. Compared with SS-DWI with the same scanning time, ESPIRiT could provide better reconstruction results with a higher SSIM of 0.56 at b = 1000 s/mm², but had artifacts indicated by an orange arrow. Furthermore, tumor aliasing was quite obvious in ResNet, as pointed by an orange arrow. The proposed method demonstrated superior reconstruction and better-defined brain tumor textures at all b-values. Then, ADC maps through different methods for the tumor case were obtained. (Figure 3 and Figure 5) Single-channel ResNet showed over-smoothness in the cerebrospinal fluid region and ESPIRiT still had structural artifacts pointed by arrows. The proposed method recovered the brain texture with the highest SSIM of 0.94.

Conclusion

We proposed a multi-channel CNN for reconstructing four-shot high-resolution DWIs from single-shot data for multiple b-values simultaneously. Cascaded with data consistency layer and ResNet, we succeeded in recovering high resolution artifact-free images using only one-fourth scanning time. Although the model is initially trained on healthy volunteers, we succeeded in inferring the pathological details of tumor patients with great generalization.

Acknowledgements

The work was supported in part by the National Natural Science Foundation of China (No. 81971583), National Key R&D Program of China (No. 2018YFC1312900), Shanghai Natural Science Foundation (No. 20ZR1406400), Shanghai Municipal Science and Technology Major Project (No.2017SHZDZX01, No.2018SHZDZX01) and ZJLab.