Improved 3D DESS MR neurography of the lumbosacral plexus with deep learning and geometric image combination reconstruction

Objective To evaluate the impact of deep learning (DL) reconstruction in enhancing image quality and nerve conspicuity in LSP MRN using DESS sequences. Additionally, a geometric image combination (GIC) method to improve DESS signals' combination was proposed. Materials and methods Adult patients undergoing 3.0 Tesla LSP MRN with DESS were prospectively enrolled. The 3D DESS echoes were separately reconstructed with and without DL and DL-GIC combined reconstructions. In a subset of patients, 3D T2-weighted short tau inversion recovery (STIR-T(2)w) sequences were also acquired. Three radiologists rated 4 image stacks ('DESS S-2', 'DESS S-2 DL', 'DESS GIC DL' and 'STIR-T(2)w DL') for bulk motion, vascular suppression, nerve fascicular architecture, and overall nerve conspicuity. Relative SNR, nerve-to-muscle, -fat, and -vessel contrast ratios were measured. Statistical analysis included ANOVA and Wilcoxon signed-rank tests. p < 0.05 was considered statistically significant. Results Forty patients (22 females; mean age = 48.6 +/- 18.5 years) were enrolled. Quantitatively, 'DESS GIC DL' demonstrated superior relative SNR (p < 0.001), while 'DESS S-2 DL' exhibited superior nerve-to-background contrast ratio (p value range: 0.002 to < 0.001). Qualitatively, DESS provided superior vascular suppression and depiction of sciatic nerve fascicular architecture but more bulk motion as compared to 'STIR-T(2)w DL'. 'DESS GIC DL' demonstrated better nerve visualization for several smaller, distal nerve segments than 'DESS S-2 DL' and 'STIR-T(2)w DL'. Conclusion Application of a DL reconstruction with geometric image combination in DESS MRN improves nerve conspicuity of the LSP, especially for its smaller branch nerves.