Quantitative $T_2^*$ Fidelity in Accelerated Radial Ultrashort-Echo-Time MRI of the Lumbar Cartilage Endplate: A Synthetic Computational Validation Study
The lumbar cartilage endplate (CEP) is a short-$T_2^*$ structure whose composition has been linked to disc health. Ultrashort-echo-time (UTE) MRI enables direct CEP visualization, but acceleration and reconstruction may alter quantitative $T_2^*$ even when image quality improves. We evaluated how radial undersampling and reconstruction strategy affect CEP $T_2^*$ fidelity in a fully synthetic setting. A 96 x 96 sagittal lumbar phantom with five CEP levels (ground-truth $T_2^*$: 24, 22, 20, 18, and 16 ms) was simulated with six UTE echoes, eight receive coils, and radial sampling. Acceleration factors $R=1,2,4,6,8$ used 302, 151, 76, 50, and 38 spokes, respectively. Reconstructions were CG-SENSE, joint multi-echo FISTA, and direct model-based estimation from k-space. Twenty independent complex-noise realizations were evaluated per acceleration. Direct model-map RMSE increased from 0.172 ms at $R=1$ to 0.800 ms at $R=4$, 1.472 ms at $R=6$, and 2.256 ms at $R=8$. At $R=4$, joint-FISTA had lower image NRMSE than the direct model (0.132 vs. 0.203), while the direct map had lower level-wise $T_2^*$ RMSE than CG-SENSE and joint-FISTA ROI fits (0.800 vs. 1.448 and 1.288 ms; estimators differed). At $R=8$, CG-SENSE ROI fitting had the lowest $T_2^*$ RMSE (1.501 ms), whereas the direct model showed systematic underestimation (bias, -2.136 ms; CCC, 0.723). Image fidelity and quantitative biomarker fidelity therefore diverged under radial acceleration. In this synthetic framework, $R=4$ was a promising operating point for direct CEP $T_2^*$ mapping, whereas $R=8$ produced substantial bias and spatial instability. Scanner, phantom, and in vivo validation are required before clinical translation.