An Accurate Estimation of T2* Mapping for Fast Magnetic Resonance Imaging

Abstract

Purpose Quantitative T2* mapping is promising due to its clinical applicability and has been employed for the assessment of cartilage repair procedures of the knee. Generally, the under-sampling k-space technique was applied to accelerate T2* mapping which is considerable time-consuming with traditional sequences. However, the under-sampling k-space technique may be impractical for acquiring reliable sampling of the T2* decay. In order to improve the corresponding accuracy of T2* mapping with under-sampled k-space technique, a new method has been proposed of deep learning (DL) based under-sampling MR images reconstruction. Methods In this work, we employed a residual network with 3 layers to explore latent functions between fully-sampled and under-sampled MR images and then applied these functions to regularize the MR images reconstruction from under-sampling k-space data. The proposed method includes three steps. Firstly, the regridding reconstruction and ESPIRiT reconstruction algorithm was used to reconstruct MR images from fully sampling and under-sampling k-space data, respectively. Then, 12 MR images at different echo time (TE=0.2/0.5/0.8/2/3.3/5.5/8/11/15/20/25/30ms) derived from under-sampling k-space data were fed into the proposed network. Ultimately, the optimized MR images were utilized to calculate T2* values. Results The T2* values derived from the proposed method were more accurate than that from the regridding reconstruction or ESPIRiT reconstruction in four tissues. For instance, when the acceleration ratio (ACC) was set at 4, the T2* values (mean ± standard deviation) of posterior cruciate ligament (PCL) were 7.97±0.40ms in regridding constructed reference image, the T2* values derived from the regridding reconstruction of 7.34±1.04ms fluctuated wildly, while the T2* values from the proposed method were restored to 7.84±1.39ms which were closer to the reference T2* values. Additionally, the T2* values of PCL were 6.99±9.47ms in ESPIRiT reconstructed reference image, the T2* values derived from the ESPIRiT reconstruction of 5.32±8.44ms fluctuated wildly, while the T2* values from the proposed method were restored to 6.64±11.73ms.The similar phenomenon can be seen in other three ROIs with ACC = 4 or 2. Conclusion T2* mapping optimized by the proposed DL-based method resembles the reference qualitatively and quantitatively. In conclusion, the proposed method has great promise on improving the accuracy of T2* mapping based on under-sampling k-space technique for fast magnetic resonance imaging.