The effect of deep learning-based compressed sensing on the image quality of contrast-enhanced 3D T1-weighted images of the maxillofacial region

Objectives: To compare the image quality of deep learning-based Compressed SENSE (DL-based CS) reconstructed images with conventional algorithm-based Compressed SENSE (Alg-based CS) reconstructed images in contrast-enhanced 3D T1-weighted images of the maxillofacial region. Methods: The cases of 32 patients who underwent two reconstructions, conventional Alg-based CS, and DL-based CS, were retrospectively analyzed. We set the reduction factors to 3 and 6. Thus, four types of images were obtained: Alg3, DL3, Alg6, and DL6. We calculated the signal-to-noise ratios (SNRs) of the muscle and spinal cord and the Structural Similarity Index Measure (SSIM) between the two reconstructions at both the level of the tongue and oral floor. Additionally, noise, visualization of anatomical structures, and motion artifacts were subjectively assessed on a 5-point scale. Results: A significant difference was observed between Alg3 and Alg6, indicating that a higher reduction factor results in a lower SNR. DL was useful for improving image quality, and no significant difference was observed between DL6 and Alg3. The SSIM at a reduction factor of six was smaller than that at a reduction factor of three, and this tendency was markedly noticeable in the lower face, which shows that DL is beneficial for denoising under low SNR conditions. The qualitative assessment of the noise and anatomical structure showed similar trends to the SNR. Conclusions: Given the quantitative and qualitative analyses, DL-based CS reconstruction is useful, especially for high reduction factors and regions with low coil sensitivity.