Introduction

Apparent diffusion coefficient (ADC) and T₂ maps have been increasingly used to cooperatively estimate various treatment modalities for diagnosis (e.g. cerebral ischemia, prostate cancer). However, the collaborative usage of ADC and T₂ maps is hampered due to the time limitation in clinical application, and also suffer from motion artifacts. Quantitative MRI methods based on overlapping-echo detachment (OLED) are able to deliver parametric maps in single-shot.^1-3 Compared to conventional image reconstruction method using separation algorithm,¹ deep-learning based method exhibits better performance in image details recovering and higher robustness to non-ideal radio-frequency (B₁) field.² In this work, we developed a single-shot synchronized diffusion and T₂ mapping based on DT₂M-OLED, which is capable of gaining reliable and self-registered diffusion and T₂ maps within 120 ms for the first time.

Methods

In the DT₂M-OLED sequence (Fig. 1(a)), six echoes with two diffusion weighting and three T₂ weighting (Fig. 1(b)) are generated by two excitation pulses α (α flip angle = 60°), refocusing pulses β (β flip angle =180°), diffusion gradient Gd and echo-shifting gradients (Gro₁, Gro₂, Gpe₁ and Gpe₂) in two echo trains.
U-net was employed for T₂ and ADC maps reconstruction. The training dataset were the real parts and imaginary parts of the overlapping-echo images from two echo trains (matrix size of 96×96) zero-filled to 160×160 and then randomly cropped into 64 × 64 matrix as the input of the network (Fig1. (c)). Totally 6000 and 1000 samples were used for training and testing, respectively. B₁ inhomogeneity and Gaussian noise were added to the training data for accounting for non-ideal experimental conditions. Adam optimizer was applied with an initial learning rate of 0.0001.
The technique was tested on numerical and rat brain experiments. Same sequence parameters were used in DT₂M-OLED acquisition for both samples. For the numerical experiment, parametric maps were also reconstructed by separation algorithm to demonstrate the advantage of deep learning reconstruction.
In-vivo experiments were conducted on a 7T Varian MRI system. Two healthy Wistar rats were used for brain scanning with DT₂M-OLED and spin-echo echo-planar imaging (SE-EPI). Resolution = 0.3 × 0.3 × 2.0 mm³. One b = 800 s/mm² for both sequences and one b = 0 image for SE-EPI. SE-EPI acquired multiple TEs (TE=50/80/110/140 ms) for T2 mapping. Reference B₁ map was calculated from EPI images with flip angle of 60° and 120°, respectively. Condition A: Rat under deep anesthetic, 16 times signal averaging for both sequences, TR = 3 s. Condition B: Anesthetic gas was closed, temporal DT₂M-OLED data were acquired every 3 s during the period of rat recovering from deep sleep without signal averaging. The acquisition time of one DT₂M-OLED scan is 114 ms. Before temporal DT₂M-OLED data acquisition, T₂ and ADC mapping by SE-EPI with and without 16-time averaging were acquired as references.

Result & Discussion

The numerical experimental results in Fig. 2 illustrate that the reconstruction based on deep learning (DL) is more robust than the one based on separation algorithm (SA). The ADC, T₂, and B₁ maps reconstructed from deep learning are in consistent with the corresponding references. DT₂M-OLED has almost the same k-space for the first echo train as DM-OLED.³ Theoretically, separation algorithm is capable to reconstruct separated images from all six echoes and deliver parametric maps. However, 1) echo3 and echo6 for B₁ estimation makes k-space more crowded. 2) The echo with diffusion weighting also has heavier T₂ weighting (echo1) than the echo without diffusion weighting (echo2) in the first echo train and echo6 with diffusion weighting owns the heaviest T₂ weighting, which intensifies the differences between overlapping echoes. These factors corrupt the quality of reconstructed images via separation algorithm. Therefore separation algorithm is not efficient enough for DT₂M-OLED reconstruction.
Fig. 3 illustrates the results of a rat brain with completely anesthetized. The ADC, B₁ and T₂ maps from DT₂M-OLED show a good concordance to those from SE-EPI in all maps. The statistic segmentation results further verify the reliability of DT₂M-OLED. The fiber orientations reflected by the ADC values are consistent with the physiological structure reported previously.⁴
Fig. 4 exhibits the results of a rat brain suffered from motion effect and lower signal-to-noise ratio (SNR). The T₂ maps show good consistency with each other while rat was moving. No signal averaging was conducted in both sequences acquisition. The parametric maps from DT₂M-OLED exhibit a bit smoother due to low SNR, but still show better contrast than SE-EPI results without signal averaging. These results indicate that DT₂M-OLED is quite robust under motion environment. Self-registration between T₂ and ADC maps makes DT₂M-OLED competent for cooperative diagnosis via T₂ and ADC mapping.

Conclusion

This study proposed an ultrafast method to deliver self-registered T₂ and ADC maps in a single acquisition, which offers the possibility of bringing joint T₂ and ADC imaging into clinical routine with unprecedented efficiency and motion robustness. Deep-learning based reconstruction further enhances the accuracy of parametric maps recovery. As a time-efficient multi-parametric imaging method, DT₂M-OLED may facilitate multi-parametric quantitative diagnosis and dynamic quantitative MRI.

Acknowledgements

This work was supported by the National Natural Science Foundation of China under grant numbers U1805261, 11761141010, 11775184 and 82071913.