This paper proposes a deep-learning-based image enhancement approach that can generate high-resolution micro-CT-like images from multidetector computed tomography (MDCT). A total of 12,500 MDCT and micro-CT image pairs were obtained from 25 vertebral specimens. Then, a pix2pixHD model was trained and evaluated using the structural similarity index measure (SSIM) and Fréchet inception distance (FID). We performed subjective assessments of the micro-CT-like images based on five aspects. Micro-CT and micro-CT-like image-derived trabecular bone microstructures were compared, and the underlying correlations were analyzed. The results showed that the pix2pixHD method (SSIM, 0.804 ± 0.037 and FID, 43.598 ± 9.108) outperformed the two control methods (pix2pix and CRN) in enhancing MDCT images (p < 0.05). According to the subjective assessment, the pix2pixHD-derived micro-CT-like images showed no significant difference from the micro-CT images in terms of contrast and shadow (p > 0.05) but demonstrated slightly lower noise, sharpness and trabecular bone texture (p < 0.05). Compared with the trabecular microstructure parameters of micro-CT images, those of pix2pixHD-derived micro-CT-like images showed no significant differences in bone volume fraction (BV/TV) (p > 0.05) and significant correlations in trabecular thickness (Tb.Th) and trabecular spacing (Tb.Sp) (Tb.Th, R = 0.90, p < 0.05; Tb.Sp, R = 0.88, p < 0.05). The proposed method can enhance the resolution of MDCT and obtain micro-CT-like images, which may provide new diagnostic criteria and a predictive basis for osteoporosis and related fractures.
Keywords: computed tomography; deep learning; osteoporosis; structure analysis; trabecular bone; vertebra.