Deep learning reconstruction enables accelerated T2-weighted MRI for rectal cancer staging: a prospective study of diagnostic consistency across NEX value reduction.
Authors
Affiliations (3)
Affiliations (3)
- West China Hospital of Sichuan University, Chengdu, China.
- GE Healthcare, MR Research, Beijing, China.
- West China Hospital of Sichuan University, Chengdu, China. [email protected].
Abstract
To evaluate the image quality, interpretation consistency, and scanning efficiency of deep learning-based reconstruction (DLR) algorithm (AIR™ Recon DL; GE Healthcare) compared with conventional reconstruction (ConR) in T2-weighted MRI for rectal cancer across different number of excitations (NEX) values. This prospective study enrolled consecutive patients undergoing MRI for primary staging of rectal cancer between July 2022 and April 2023. Each patient underwent T2-weighted MRI with three NEX values (4, 2, and 1), reconstructed using both ConR and DLR methods, generating six image sets. Image quality was assessed quantitatively (signal-to-noise ratio, contrast-to-noise ratio) and qualitatively using a 5-point scale for tissue contrast, edge sharpness, rectal wall morphology, tumor characteristics, overall image quality, and noise. Two radiologists independently evaluated staging parameters including T stage, N stage, extramural vascular invasion (EMVI), and mesorectal fascia (MRF) status. Inter-observer and inter-sequence agreement were analyzed. Thirty-five patients completed all six imaging protocols. Acquisition times were 3 min 55s (NEX = 4), 2 min 1s (NEX = 2, 49% reduction), and 1 min 4s (NEX = 1, 73% reduction). For qualitative assessment, DLR sequences demonstrated significantly superior image quality compared to ConR sequences across all NEX values (all P < 0.001). Among all sequences, FSE<sub>NEX2-DLR</sub> achieved the highest qualitative scores for tissue contrast, edge sharpness, tumor characteristics, and overall image quality. Quantitative analysis showed FSE<sub>NEX4-DLR</sub> demonstrated the highest SNR and CNR values, followed by FSE<sub>NEX2-DLR</sub>, with all DLR sequences significantly outperforming their ConR counterparts (all P < 0.001). Inter-sequence agreement analysis revealed substantial to almost perfect consistency for all staging parameters (mean κ: 0.736-0.800). DLR sequences demonstrated significantly higher inter-sequence agreement (range: 0.715-0.964) compared to ConR sequences (range: 0.505-0.811). DLR with reduced NEX values maintains diagnostic consistency in rectal cancer staging. FSE<sub>NEX2-DLR</sub>, achieving 49% scanning time reduction while preserving diagnostic consistency, may offer a reasonable balance between acquisition efficiency and image quality. Future studies with histopathological correlation are needed to validate diagnostic accuracy.