INTRODUCTION

Simultaneous measurement of cerebral blood flow (CBF) and apparent diffusion constant (ADC) could provide important diagnostic and prognostic information for patients¹. Recent studies^2-3 have indicated that the directional sensitivity of the intravoxel incoherent motion (IVIM) signal to flow could characterize the direction-dependent flow network in cerebral vasculature and assess ischemic infarction and perfusion in patients with acute ischemic stroke. Generally, IVIM can measure both water molecule diffusivity and perfusion information with a variable range of b values using a double-exponential model. The IVIM signal is often averaged over a region of interest (ROI) to improve SNR and stability to avoid overfitting problems. However, the sensitivity of ROI-based approaches depends on the visual inspection of the ROIs. Recent researches^4-5 showed that diffusion weighted images with higher resolution could provide the higher sensitivity of detecting ischemic lesions in TIA patients.
This work aimed to evaluate the performance of a deep learning-based high resolution multi-shot IVIM parameter mapping method in assessment of lacunar infarction patients.

METHODS

MR Imaging: All MRI experiments were performed on a Philips 3.0 T clinical scanner (Philips Healthcare, Best, the Netherlands). Brain images were acquired from 17 patients (mean age, 58.4 years; range, 48~71 years; 9F/8M) with lacunar infarction using a 32-channel phase-array head coil. All DWI data were acquired with 4-shot interleaved multi-shot (MSH) EPI DWI sequence from 20 slices, covering the whole brain with slice gap of 6 mm. Other parameters were shown as following: FOV=228×252 mm², acquisition matrix=228×252, special resolution=1×1×5 mm³, TR/TE=2800/65 ms, scan time ≈400 s, b values = 0, 20, 40, 60, 80, 100, 120, 200, 400, 600, 800 and 1000 s/mm² without additional average. Single-shot (SSH) EPI DWI were obtained with the same ETL and diffusion weighting protocols.
Deep learning reconstruction: We trained the network on SSH-EPI data and applied the model to MSH-EPI. The aliased images were generated by down-sampling the full k-space with the same under-sampling factor and trajectory as MSH-EPI. Dual-channel U-net including multiple convolutional layers is used for phase estimation of MSH-EPI. The schematic diagram of deep learning based MUSE⁶ is shown in Fig. 1. Similar to conventional MUSE, images were reconstructed by incorporating the corrected phase into the final reconstruction. Instead of using SENSE for phase estimation, our method is performed by U-net with k-space data from different shots of DWI.
By incorporating U-net into conventional MSH-EPI reconstruction method⁷, the efficiency and performances of MSH-EPI diffusion weighted images could be improved. Meanwhile, the reconstruction of MSH-EPI using conventional model-base method and SSH-EPI were compared based on ROIs.
Image analysis: The D_s and D_f maps were calculated from the whole brain DW images using a least square nonlinear fitting algorithm in MATLAB (The Mathworks Inc., Natick, MA) according to the following equation. 1:
S=S₀*(f*exp(-b*D_f) + (1-f)*exp(-b*D_S)) [1]
S=S₀*exp(-b*ADC) [2]
where D_s is diffusion coefficient, D_f is perfusion-related pseudo-diffusion coefficient and f is perfusion fraction. And ADC maps were fitted pixel-wisely from the whole brain according to equation.2.

RESULTS

Compared with conventional reconstruction, deep-learning based high resolution DWI showed high advantage in alleviating motion induced artifacts, shown in Fig. 2 (black arrows showed the residual aliasing artifacts in phase results) and Fig.3 (red arrows showed the residual aliasing artifacts in diffusion weighting images) for different b-values (shown with representative high b-values: 400, 600, 800 and 1000 s/mm²) on a patient with old cerebral lacunar infarction. In quantitative comparison, the ADC of deep-learning based method showed clearer boundary for malacia (black arrows in ADC maps of Fig.4 and white arrows in the corresponding D_S maps of Fig.4) and infarction lesions (black arrows in Fig.5) compared with conventional method.

CONCLUSION

The proposed deep-learning based high resolution IVIM method can qualitatively and quantitatively improve the performance.

Acknowledgements

This work was supported by Shanghai Municipal Science and Technology Major Project (No.2017SHZDZX01), Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01) and ZJLab, Shanghai Natural Science Foundation (No. 17ZR1401600) and the National Natural Science Foundation of China (No. 81971583).