Deep learning reconstruction enhances bone visualization in zero echo time MRI for cervical spondylosis: A prospective study

Abstract

Objective

To evaluate the use of deep learning-based reconstruction (DLR) in zero echo time (ZTE) MRI to improve image quality and reduce scan time for assessing cervical spondylosis.

Methods

Forty-three preoperative patients with cervical spondylosis underwent ZTE MRI. ZTE data were acquired with two different numbers of excitations (NEX): NEX3 and NEX8. In addition to the conventional reconstruction, the NEX3 images were also reconstructed using DLR (NEX3-DL). Quantitative comparisons of signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) across cortical bone, cancellous bone, spinal canal, and posterior muscles were analyzed via Wilcoxon signed-rank tests. Additionally, non-inferiority testing (NEX3-DL vs. NEX8) and superiority testing (NEX3-DL vs. NEX3) were performed. Two groups of musculoskeletal radiologists independently delineated regions of interest and evaluated the following parameters using validated assessment scales: cortical bone depiction, artifacts, perceived image noise, overall image quality. Inter- and intra-reader agreement were assessed with κ values and intraclass correlation coefficients (ICC).

Results

NEX3-DL images demonstrated significantly improved SNR and CNR compared to NEX3, matching the performance of NEX8 images with a 62 % shorter scan time. Subjective quality scores of both NEX3-DL and NEX8 were superior to NEX3 (p < 0.01) with no significant differences between them. Inter- and intra-rater reliability for continuous variables and ordinal assessments showed substantial to excellent agreement (κ or ICC > 0.6).

Conclusion

DLR-enhanced ZTE MRI improves bone visualization and can be integrated into routine MRI protocols, facilitating comprehensive evaluation of osseous and soft tissue structures within a single, time-efficient, radiation-free examination, and streamlining clinical workflow.