Synopsis

Keywords: Quantitative Imaging, Machine Learning/Artificial Intelligence

A Deep learning (DL)-based reconstruction is a promising method to achieve higher resolution for diffusion-weighted Kurtosis imaging (DKI) without increasing signal averaging. The DKI phantom and patient results demonstrated improved image quality and reduced Gibbs (ringing) artifact, aiding in the robust estimation of D_app and K_app. In all phantom and patient data, the standard deviation of D_app and K_app measured in images reconstructed without DL was higher than in images reconstructed using DL. The NEX=1 significantly reduced the multi-b-value data acquisition time, and the DL-based reconstruction can produce images comparable to the standard NEX=2 or 4, depending on the b-value.

Purpose:

Non-Gaussian diffusion kurtosis imaging (DKI) measures quantitative metrics describing the water diffusivity of hindered and restricted water molecules and tissue microstructure^1,2. For extra-cranial regions such as the head and neck (HN), multi-b-value data acquisition is challenging, particularly at higher b-value with a reduced number of excitations, to maintain the optimal signal-to-noise ratio (SNR) required for DKI analysis^2,3. A linear fit to the natural logarithm of the multiple b-value data with DKI modeling yields the apparent diffusion (D_app) and kurtosis coefficient (K_app), which are the surrogate markers of the tumor cellularity and tissue microstructure². A recently developed novel deep learning (DL)-based MR reconstruction method (AIR^TM Recon DL) has enhanced the image signal noise ratio (SNR), sharpness, and reduced truncation artifacts in prostate cancer patients^4-6. Tumors in the HN region include a diverse group of cancers, and the accurate measurement of quantitative metrics may improve treatment response assessment for such tumors^7-9. This study investigates whether DL-based reconstruction can improve the quantification of DKI metrics for tumors in the HN region.

Methods:

MRI Data Acquisition: Data were acquired on a 3T MRI scanner (SIGNA Premier, GE Healthcare) using a 21-channel HN unit.
Phantom: The novel, multi-exponential DKI phantom used in this study has two tiers of seven 20ml-glass scintillation vials containing lamellar vesicle (LV) solutions placed in a 1L jar filled with deionized water (Figure 2)^10,11. The LV materials were made from fatty alcohols, surfactants, and water (solid-in-water %(w/w)=0.5%-2.5%). Two polyvinylpyrrolidone (PVP) at 20% and 40% were added as Gaussian-diffusion (K_app=0) “controls.” The phantom contained an alcohol thermometer for temperature (T) reading to ±0.5°C. The phantoms were left to thermolyze in scanner rooms overnight before scanning, and the temperature during the exam was 19°C. The multi b-value DW images were acquired using a single shot spin echo planar imaging (SS-SE-EPI) sequence with TR/TE=4000/66 (minimum) ms, the field of view (FOV)=20 cm, matrix=128×128, slices=15, slice thickness=5mm, and b=0,20,50,80,200,300,500,1000,1500,2000 s/mm². Two sets of data acquisition were performed using the above parameters with modification in the number of excitations (NEX): 1) Using the standard NEX=2 for b≤300 and NEX=4 for b≥500, and 2) Reducing the NEX=1 for all the b-values. The total scan time with the standard NEX was 6.12 min and for reduced NEX=1 was 2.13 min.
Patients: DW-MRI data were acquired from six HN cancer patients (median age 59 years, six males, 2 HPV(+) positive, 1 HPV(-), and 3 with unknown primary tumor status) in this retrospective study between December 2021 and June 2022. These patients underwent chemo-radiation therapy (CRT), and all MRI examinations were performed before treatment. MRI protocol consisted of multi-planar T1/T2 weighted imaging followed by multi-b-value DWI (b-values same as mentioned above) with TR/TE=4000/80 (minimum) ms, the field of view (FOV)=20-24 cm, matrix=128×128, slices=8-10, slice thickness=5 mm, number of excitation (NEX)=2 and b=0,20,50,80,200,300,500,1000,1500,2000 s/mm². The raw data from the DW-MRI scans were transferred and retro-reconstructed using the AIR^TM Recon DL algorithm in the GE reconstruction pipeline (Orchestra SDK, GE Healthcare) and finally labeled as DL images ^4,5.
ROI Contouring and DWI Data Analysis: Regions of Interest (ROIs) were delineated on the primary tumors and neck nodal metastases by an experienced neuro-radiologist on both DW images with and without DL (b = 0 s/mm²) using ITK-SNAP. All DW data analysis was performed using in-house software, MRI-QAMPER (Quantitative Analysis Multi-Parametric Evaluation Routines), written in MATLAB (MathWorks, Natick, MA). ROI analysis yielded mean and standard deviation (STD) for each reported metric. Coefficient of variation (CV) and relative percent of difference reported.

Results:

The DL-based reconstructed images obtained from the DKI phantom showed reduced Gibbs (ringing) artifacts and increased image sharpness (Figure 1). Figure 2 exhibits the D_app and K_app maps for the standard and reduced NEX images reconstructed with and without DL for the DKI phantom. D_app demonstrated no significant difference between the standard and reduced NEX images reconstructed using DL, while K_app has shown a difference between the two acquisition and reconstruction strategies, particularly for vials #1, #6, and #7. The result of this comparison in the DKI phantom can be observed in Figure 3. Figure 4 shows data from a representative patient with primary tumor and neck nodal metastases. The improved SNR and reduced artifacts can be observed in the patient data, similar to phantom results (Figure 4). Table 1 summarizes the mean and CV of D_app and K_app in 10 ROIs for images with and without DL. For mean D_app, a maximum difference of 5.22% was observed between the metrics derived from the images with and without DL, and for the mean K_app, a maximum difference of 11.36% was observed between these two sets of images.

Discussion and Conclusion:

The DL-based image reconstruction improved DKI image quality with increased sharpness and reduced Gibbs (ringing) artifacts. In all phantom and patient data, the STDs of D_app and K_app measured in images reconstructed without DL were higher than in images reconstructed using DL. Particularly at higher b-values, the DL-based image reconstruction improved image quality, reducing bias in estimated D_app and K_app. The NEX=1 significantly reduced the multi-b-value data acquisition time, and the DL-based reconstruction can produce images comparable to the standard NEX=2 or 4, depending on the b-value.

Acknowledgements

Funding support from National Institutes of Health Grant: U01 CA211205 (ASD)