Deep learning-based k-space-to-image reconstruction and super resolution for diffusion-weighted imaging in whole-spine MRI

Purpose: To assess the feasibility of deep learning (DL)-based k-space-to-image reconstruction and super resolution for whole-spine diffusion-weighted imaging (DWI).

Method: This retrospective study included 97 consecutive patients with hematologic and/or oncologic diseases who underwent DL-processed whole-spine MRI from July 2022 to March 2023. For each patient, conventional (CONV) axial single-shot echo-planar DWI (b = 50, 800 s/mm²) was performed, followed by DL reconstruction and super resolution processing. The presence of malignant lesions and qualitative (overall image quality and diagnostic confidence) and quantitative (nonuniformity [NU], lesion contrast, signal-to-noise ratio [SNR], contrast-to-noise ratio [CNR], and ADC values) parameters were assessed for DL and CONV DWI.

Results: Ultimately, 67 patients (mean age, 63.0 years; 35 females) were analyzed. The proportions of vertebrae with malignant lesions for both protocols were not significantly different (P: [0.55-0.99]). The overall image quality and diagnostic confidence scores were higher for DL DWI (all P ≤ 0.002) than CONV DWI. The NU, lesion contrast, SNR, and CNR of each vertebral segment (P ≤ 0.04) but not the NU of the sacral segment (P = 0.51) showed significant differences between protocols. For DL DWI, the NU was lower, and lesion contrast, SNR, and CNR were higher than those of CONV DWI (median values of all segments; 19.8 vs. 22.2, 5.4 vs. 4.3, 7.3 vs. 5.5, and 0.8 vs. 0.7). Mean ADC values of the lesions did not significantly differ between the protocols (P: [0.16-0.89]).

Conclusions: DL reconstruction can improve the image quality of whole-spine diffusion imaging.