Prospective Validation of Accelerated Brain MRI Using Deep Learning-Based Reconstruction: Simultaneous Application to Spin-Echo and Gradient-Echo Sequences

KCR (2023)

PURPOSE

To evaluate the effectiveness of brain MRI with deep learning-based fast reconstructed MRI (DL-FR) by comparing qualitative and quantitative image quality of conventional MRI for both spin-echo and gradient-echo sequences.

MATERIALS AND METHOD

In this multi-reader, multi-sequence, and multi-vendor prospective study, both synthetic and conventional MRI pairs with T1-weighted, T2-weighted, T2-FLAIR, and 3D T1-weighted images from 100 subjects (29 males; 57.3±16.0 years; 34 subjects with GE, 35 with Philips, and 31 with Siemens), were examined by 4 blinded neuroradiologists. The images were randomized, and evaluated for overall image quality, delineation of structures, and artifacts using 5-point Likert scale, in the two separate sessions. Intraclass correlation coefficient was calculated for Inter-reader agreement. Signal-to-noise ratio (SNR) and contras-to-noise ratio (CNR) was obtained for both 3D T1w and T2 FLAIR. For quantitative analysis, volumetric analysis of brain parcellation and white matter T2 hyperintensity (WMH) was performed using LesionQuant (LQ) and NeuroQuant (NQ), respectively.

RESULTS

DL-FR showed median 41% (24-51%) reduced acquisition time with improvements in overall image quality (mean±SD, 3.79±0.72 vs. 3.40±0.63, p<0.001); and enhanced structure delineation (3.59±0.81 vs. 3.45±0.77, p<0.001), while lesion conspicuity (3.34±1.05 vs. 3.26±1.08, p=0.32), Fazekas scale (2.24±1.40 vs 2.18±1.40, p=0.37), enlarged perivascular space grading (1.93±1.21 vs 1.90±1.20, p=0.65), and image artifacts (3.78±0.73 vs. 3.79±0.69, p=0.47) was comparable. Inter-reader agreement was moderate to substantial (κ=0.74 for structure and lesion delineation; κ=0.52 for artifact; and κ=0.55 for overall quality), indicating reliable qualitative assessments. SNR and CNR was increased compared to conventional MRI (82.0±23.1 vs 31.4±10.8, p=0.02; and 12.4±4.1 vs 4.4±11.2, p=0.02). In the volumetric analysis, no significant differences were observed in 1253 out of 1276 (98.2%) regional volumes (r=0.92±0.10, range 0.38-1.00), with the exception of the deep gray matter including thalamus. Additionally, 5 out of 6 (83.3%) lesion categories showed no significant differences, except for juxtacortical lesion counts (r=0.64±0.29, range 0.12-0.92).

CONCLUSION

Deep learning-based fast reconstruction significantly improves the efficiency and quality of brain MRI in both spin-echo and gradient-echo sequences without compromising lesion detection and quantitative volumetric analysis.