Fast and High‐Resolution T2 Mapping Based on Echo Merging Plus k‐t Undersampling with Reduced Refocusing Flip Angles (TEMPURA) as Methods for Human Renal MRI

To develop a highly accelerated multi‐echo spin‐echo method, TEMPURA, for reducing the acquisition time and/or increasing spatial resolution for kidney T2 mapping.


INTRODUCTION
MR T 2 relaxation time mapping can provide quantitative T 2 measurements that are independent of confounding factors related to imaging sequences and hardware.This provides a more precise and reproducible method for evaluating the severity of pathological changes compared to conventional T 2 -weighted MRI.T 2 values are indicative of tissue composition, particularly free water content, and are sensitive to tissue hydration or edema.In renal imaging, T 2 mapping has shown its potential in evaluating several diseases including autosomal dominant polycystic kidney disease (ADPKD), 1 ischemia-reperfusion injury, 2,3 renal transplants, 4,5 and renal cell carcinoma. 6owever, routine clinical T 2 mapping has been restricted by the typically long acquisition times, particularly in the abdomen with respiratory triggering.Moreover, the length of time restricts spatial resolution and therefore its ability to probe tissue heterogeneity and tumor habitats, which in turn affects its accuracy in characterizing pathological changes. 7For instance, the standardized cross-vendor 2D multi-echo spin-echo (MESE) sequence developed for the UK Renal Imaging Network MRI Acquisition and Processing Standardization (UKRIN-MAPS) project [8][9][10] takes approximately 4 min for a 3 × 3 mm 2 in-plane acquisition, making higher spatial resolution acquisitions impractical for clinical use.
This study introduces a highly accelerated MESE method, termed T 2 mapping using Echo Merging Plus k-t Undersampling with Reduced refocusing flip Angles (TEMPURA).A fast breath-hold sequence and a high spatial resolution sequence were both implemented based on TEMPURA.Their performance was compared with the standardized UKRIN-MAPS sequence and a fast sequence accelerated by purely k-t undersampling in both phantom and in vivo experiments studying the kidney.

Acquisition schemes
Figure 1 illustrates the different MESE acquisition schemes investigated in this study.The standardized cross-vendor UKRIN-MAPS sequence is respiratory-triggered and uses a SENSE factor of 3 (Standardized, Figure 1A).To further exploit sparsity, data can be randomly undersampled in both the ky and echo dimensions and reconstructed based on the CS theory (k-t CS, Figure 1B).The TEMPURA schemes combine every three adjacent echoes into one k-space, either by combining three independent echoes (echo-combination, Figure 1C) or sharing one echo between two k-spaces (echo-sharing, Figure 1D).The combined k-space is still sparsely sampled.Smaller flip angles (175

Reconstruction of undersampled k-space
For the undersampling reconstruction of k-t CS and TEMPURA, we adopted a CS-based method using a self-calibrating temporal principal component analysis (PCA) basis for transform sparsity.Initial images (m 1 ) were firstly reconstructed from undersampled k-space data (k) by the k-t FOCUSS approach, 25 which were then used to estimate multicoil bases (U) for PCA based on eigen decomposition process 11 : where W is the weight matrix of k-t FOCUSS, and F t and F are fast Fourier transform along the temporal and spatial dimensions, respectively.
The final images (m 2 ) were then updated by soft-thresholding using the estimated bases: where coil sensitivity (S) was estimated from a compact calibration region of k-space data using the ESPIRiT method. 26The regularization parameters  1 and  2 were optimized and both set to 0.001 in the study.

Fitting using the extended phase graph model
To address the indirect echoes resulting from reduced refocusing flip angles, we use an extended phase graph (EPG) model, 27 adapted from the StimFit toolbox, 28,29 for fitting in this study: where T 2 can be calculated based on the known values of T 1 , refocusing flip angle (FA), echo spacing (ESP), echo train length (ETL), and signal intensity of each echo S i .Considering that the EPG model is not sensitive to T 1 variation, 28 a fixed T 1 of 1500 ms was used in all processing.For TEMPURA, a fixed refocusing FA of 110 • without scaling of transmit field B 1 was used due to challenges in fitting B 1 variation within the echo merging scheme.

Imaging protocol and implementation
Two versions of the TEMPURA sequence were developed: one highly accelerated single breath-hold sequence and one high-resolution respiratory-triggered sequence using a larger matrix size.Parameters for each of the sequences are shown in Table 1.
All experiments were performed on a 3T scanner (Discovery MR750; GE Healthcare, Waukesha, WI) with a 32-channel cardiac array coil.Sequence and reconstruction parameters were optimized based on measurements from the ISMRM/NIST phantom. 30Original T 2 -weighted images were reconstructed from undersampled data after echo emerging in k-space, then fitted using the EPG model based on the StimFit toolbox. 28,29All processing was performed offline using MATLAB (MathWorks, Natick MA).
The average computation times were 2.2 and 7.8 min for breath-holding and high-resolution acquisitions (Intel i9-13900KF, 64 GB RAM).

Phantom experiments
The ISMRM/NIST system phantom 30 was used to evaluate the accuracy of T 2 measurements against temperature-corrected reference values. 31The plate containing 13 T 2 spheres filled with MnCl 2 -doped water and a resolution inset was scanned using previously described sequences, together with the single-slice fully sampled NIST reference sequence (see Table 1).Two studies were conducted to assess the impact of TEMPURA on (1) acceleration and (2) spatial resolution.The acceleration evaluation involved acquisitions with a fixed matrix size (128 × 128) but varying the acceleration factor (×1, ×3.3, ×6.5, ×9.4, and ×11.3), corresponding to acquisition times of 2:40, 0:51, 0:26, 0:18 and 0:15 (min:s).The TR was reduced to 1125 ms to further reduce the acquisition time.For the spatial resolution evaluation, sequences had a fixed acquisition time (2:49, matched to the standardized UKRIN acquisition) but were collected at seven different matrix sizes (128 × 128, 192 × 192, 256 × 256, 320 × 320, 384 × 384, 448 × 448, and 512 × 512), resulting in in-plane spatial resolutions from 3.00 to 0.75 mm.Each sequence was repeated three times.

In vivo experiments
The kidneys of 16 healthy subjects (9 men; 7 women; range 24-47 y), one patient with a renal oncocytoma (male, 74 y), and one patient with clear cell renal cell carcinoma (ccRCC) (male, 62 y) were prospectively imaged.
Studies were approved by the local research ethics committee, and all participants gave informed consent.The standardized UKRIN sequence, breath-hold TEMPURA sequence with echo-sharing, k-t CS, and the high-resolution TEMPURA sequence with echosharing (384 × 384) were collected on each subject.Synthetic T 2 -weighted images were generated from the high-resolution T 2 and M 0 maps without additional acquisitions.A separate T 2 -weighted 3D fast spin echo (FSE) sequence was also collected on the patient (respiratorytriggered, FOV 400 × 360 × 192 mm, matrix 256 × 224 × 48, TE/TR 67.9/8574 ms, echo train length 120).
Regions of interest (ROIs) were manually drawn on the standardized UKRIN T 2 maps to define the whole kidney, and renal cortex and medulla.Minor manual adjustments were made to correct for motion to apply these ROIs to the T 2 maps generated for the other sequences.The mean T 2 values from the cortex, medulla and whole kidney were measured.

Statistical analysis
In phantom experiments, accuracy of T 2 measurements was assessed by calculating mean absolute percentage error (MAPE) and pixelwise RMS error (RMSE) against reference values (T ref 2 ) for seven spheres within the physiologically relevant range (42-405 ms).
MAPE assesses the overall bias between the averaged T 2 values of each sphere and the reference values: RMSE compares all pixels in selected spheres with the reference values on a pixel-by-pixel basis: SNR was measured to evaluate the image quality of the original T 2 -weighted images.The noise level was estimated by placing an ROI on deionized water filling, which has a long T 2 , and calculating the standard deviation of signals across both the ROI and echoes.The signal level was estimated by averaging the signal intensity across the entire phantom.
For phantom experiments, all repeated measurements were compared with the phantom reference values using a random-intercept linear mixed-effects model (details in Appendix A).For in vivo experiments, the results of all methods were compared with the reference measurements obtained by the standardized UKRIN method using MAPE, a paired Student's t-test, Pearson correlation analysis and Bland-Altman analysis.P values <0.05 were considered to be statistically significant in all analyses.

RESULTS
The results of the T 2 measurement in the ISMRM/NIST system phantom are presented in Figure 2. Figure 2A,B show the MAPE, RMSE pixel , and SNR for the different acceleration factors to collect images at a given 3 mm in-plane spatial resolution, and for different matrix sizes in a given acquisition time, respectively.Figure 2C shows the regression plots of the T 2 measurements in each sphere against reference values across selected acquisitions.Among the three acceleration methods, TEM-PURA with echo-sharing demonstrates the highest accuracy.Compared to the standardized method, the acquisition time can be reduced from 169 to 18 s, with the MAPE of TEMPURA (8.2% at 9.4×) remaining comparable to the standardized UKRIN method (7.4%) but much lower than that of k-t CS (21.8% at 9.4×).TEM-PURA also outperformed k-t CS in RMSE pixel (TEM-PURA 9.4×: 19.9, k-t CS 9.4×: 36.9, standardized: 12.9) and SNR (TEMPURA 9.4×: 13.5, k-t CS 9.4×: 11.9, standardized: 16.1).
Employing larger matrix sizes to increase the spatial resolution in TEMPURA echo-sharing greatly improved the visualization of the detailed structure without increasing the acquisition time.High-resolution TEMPURA exhibits reduced MAPE values (4.8% and 6.1% for 384 × 384 and 512 × 512) compared to the standardized method, possibly due to a larger number of samples and less partial volume effect resulting from increased resolution.The RMSE pixel and SNR of high-resolution TEM-PURA are similar to the standardized method.
Figure 3 shows representative results from a healthy volunteer, a patient with an oncocytoma, and a patient with a ccRCC.In Figure 3A, single breath-hold TEMPURA produced comparable image quality to the standardized UKRIN MESE sequence, whereas k-t CS led to image blurring.High-resolution TEMPURA with a 3× matrix size substantially enhanced the imaged anatomical detail in the kidney, allowing the cortex and medulla to be distinguished on the T 2 map.In Figure 3B,C, while breath-hold TEMPRUA yields T 2 maps similar to standard MESE, high-resolution TEMPURA improves the tumor visualization in patient images, enabling clear observation of its detailed structure and T 2 distribution of the habitats within the tumor.The synthetic T 2 -weighted image shows superior image quality and anatomical detail compared to the T 2 -weighted images acquired by a separate 3D FSE sequence.
In vivo T 2 measurements from the volunteers are shown in Table 2 and Bland-Altman plots in Figure S1.Using the standardized UKRIN method as the reference, both breath-hold and high-resolution TEMPURA echo-sharing achieved good agreement (MAPE = 1.31%-2.50%and 2.80%-3.28%,respectively) and high correlation coefficient (R = 0.85-0.98 and 0.82-0.96,respectively; p < 0.001), whereas k-t CS showed a much lower correlation (0.57-0.59, p < 0.05) and higher Quantitative evaluation of T 2 measurements from the NIST phantom.MAPE (3.28%-4.45%).No significant difference was found between each method and the standardized method except the T 2 measurements in medulla for k-t CS and high-resolution TEMPURA.
The signal evolution and fitting curves of TEMPURA and standard MESE are displayed in Figure S2.TEMPURA exhibited a slower signal decay due to using smaller refocusing FAs, yet both methods yielded similar T 2 measurements.The standard MESE shows zigzag curves caused by B 1 inhomogeneity, whereas TEMPURA with echo-sharing fills even echoes into the inner portion of k-space and results in smoother signal evolution curves.

F I G U R E 3
Representative images from a healthy volunteer (A), a patient (male, 73 y) diagnosed with an oncocytoma (B) and a patient (male, 62 y) diagnosed with ccRCC (C).In (A), breath-hold TEMPURA produced similar image quality as the standardized method, while k-t CS resulted in blurred images.High-resolution TEMPURA enhanced the anatomical detail, allowing differentiation between the cortex and medulla.In (B) and (C), high-resolution TEMPURA enabled clear observation of its detailed structure and T 2 distribution.The upper row shows zoomed-in views of the tumor region.The synthetic T 2 -weighted image (TE = 162 ms) also shows greater anatomical detail compared to the T 2 -weighted images acquired by a separate 3D FSE sequence.

DISCUSSION
We have presented TEMPURA, a highly accelerated and high spatial resolution method for T 2 mapping based on k-t undersampling and echo merging.Reduced refocusing FAs allowed for more echoes within SAR limits.TEM-PURA outperformed the k-t CS method that simply undersamples k-space, maintaining low errors and no significant difference from the reference T 2 values when the acceleration factor was no greater than 9.4×.The high time efficiency enables a fast breath-holding acquisition, greatly reducing the acquisition time of renal T 2 mapping from approximately 4 min to a single breath-hold of 18 s.Moreover, the acceleration of TEMPURA was utilized to increase the spatial resolution without prolonging the acquisition time, enhancing visualization of detailed anatomy within the kidney and facilitating the improved differentiation of renal cortex and medulla.The improved resolution also allowed the identification of intratumoral heterogeneity and tissue habitats within a renal mass, which could benefit tumor stratification and texture analysis.Synthetic T 2 -weighted images across variable TEs were generated without additional acquisitions: compared with the 3D FSE sequence, synthetic T 2 -weighted images provided a clearer depiction of the renal anatomy and the intratumoral heterogeneity due to the higher resolution and less T 2 blurring caused by mixing k-space data from different echoes. 23This high-resolution T 2 mapping approach, along with the synthetic T 2 -weighted images, may be used for investigating disease-related changes in both morphology and quantitative T 2 values.
In this study, we demonstrated that echo-sharing is more accurate than echo-combination for merging echoes in TEMPURA, possibly due to two factors.First, echo-sharing allocates all even echoes into the inner portion of k-space, and even echoes are more accurate than odd echoes for T 2 quantification in MESE. 32Second, the echo sharing approach allows for smaller ΔTE and more echoes, which is advantageous when measuring small T 2 components.
While we set the number of combined echoes to three in this study, it may be further increased for acquisitions involving more echoes.Multi-compartment T 2 acquisitions based on MESE are particularly suitable for TEMUPRA due to their long TEs and large echo numbers.This allows for combining more echoes and using higher k-t undersampling factors.Future studies will investigate the application of TEMPURA in multi-compartment T 2 acquisitions, such as luminal water imaging in the prostate. 33EMPURA achieves a 9.4-fold acceleration, surpassing previous k-t CS models (2-to 3-fold in ( 12) and 4-to 8-folds in ( 13)) and deep-learning methods using purely undersampling (8-fold in ( 14) and ( 15)).While GraSE presents as another rapid sequence, 17,18 its utilization of gradient echoes with lower SNR and T 2 * weighting may cause T 2 overestimation, 18 image blurring, 34 as well as peripheral nerve stimulation and acoustic noise. 34TEMPURA merges spin echoes with small echo spacings, potentially alleviating these issues.MARTINI 22 and GRAPPATINI 23 also merge adjacent echoes of MESE but lack CS reconstruction or reduced FAs.While MATINI/-GRAPPATINI showed a 10-fold acceleration, they were only used for acquisitions with large matrix sizes (260 × 320-512 × 270) and long acquisition times (2:50-6:22), unlike TEMPURA's more challenging single breath-hold acquisitions.Future studies should comprehensively compare TEMPURA with other methods.
This study has several limitations.First, the current fitting method is a simplified EPG-based model that does not fully consider the echo merging scheme and B 1 inhomogeneity, which will be improved in future work.Second, manual placing ROIs on standardized T 2 maps for the cortex and medulla might compromise accuracy due to the difficulty in distinguishing these regions on low-resolution T 2 maps.Furthermore, image misregistration between T 2 maps acquired by different methods may not be fully corrected by manual adjustment in this study, particularly between breath-hold and respiratory triggered methods.With a more robust image registration method, signal heterogeneity analysis and comparison can also be performed.Third, a repeatability evaluation comparing the in vivo scan-rescan variability of TEMPURA with other methods should be performed in future work.Fourth, to address potential challenges associated with offline processing in clinical implementation, efforts should focus on developing faster online reconstruction and fitting methods using high-speed programming languages.Last, we have only demonstrated this method for renal T 2 mapping, and only two patients were imaged as part of this study.TEMUPRA has potential to be applied in various other anatomical regions.The breath-hold version could be particularly beneficial for other organs that experience respiratory motion, such as the liver, spleen, pancreas and cardiac measurement.Future research will investigate the performance of TEM-PURA in larger cohorts of patients and across other body regions.

CONCLUSIONS
We have developed single breath-hold and high spatial resolution renal T 2 mapping sequences using a new acceleration method termed TEMPURA.The breath-hold sequence offers a rapid and accurate T 2 measurement, which can be potentially used for the diagnosis of renal diseases requiring quick examinations.The high-resolution sequence provides the distinct depiction of anatomical structures within the kidney, facilitating in depth evaluation of both anatomical morphology and quantitative T 2 values for a diverse range of pathological conditions including intratumoral heterogeneity.

F I G U R E 1
Schematic diagram of 2D MESE pulse sequences with different acceleration methods.
(A) MAPE, RMSE pixel , and SNR of T 2 measurements using different acceleration methods with different acquisition times.The matrix size is 128 × 128 for all acquisitions.(B) MAPE, RMSE pixel , and SNR of T 2 measurements using TEMPURA (echo-sharing) with different matrix sizes.The acquisition time is 169 s for all TEMPURA acquisitions and the standardized acquisition.(C) Regression plots indicating measurements from T 2 spheres in comparison with the NIST reference values.Red rectangular boxes indicate physiologically meaningful ranges (45-500 ms).Upper images show the resolution inset (indicated by the yellow arrow on the phantom image) cropped from corresponding source images (TE = ∼40 ms).
Key parameters of the different MESE-based T 2 mapping sequences.
T A B L E 1Note: Other parameters in common: FOV = 384 mm, five slices with thickness/gap of 4.5/1.0mm (except NIST: FOV = 250 mm, one slice with thickness 6 mm).
Comparison of T 2 measurements from 16 healthy volunteers using TEMPURA and k-t CS versus the standardized method.MAPE, Pearson correlation coefficients, and paired Student's t-tests are calculated between each method and the standardized method.