Deep learning model trained using multi-energy computed tomography (CT) data shows better metal artifact reduction for lumbar CT imaging

AIM:

To develop different deep learning–based metal artifact reduction (MAR) models (deep-MAR) based on virtual monochromatic images (VMIs) at both multiple energy levels and single-energy level, and compare their performance under wider energy levels.

MATERIALS AND METHODS

We enrolled 93 patients with lumbar implants who underwent multi-energy CT scans and then reconstructed into VMIs at energy levels of 70 and 100 keV (10 randomly selected cases at levels ranging from 40 to 140 keV). Original images processed by modelMAR were served as the established reference. Deep-MAR models were trained using diverse datasets at energy levels of 70 KeV, 100 KeV, and two levels (model70, model100, and modelmix). Afterwards, original images were processed using three deep-MAR models, and peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were calculated. Objective and subjective image qualities of all processed images were also compared.

RESULTS

From 40 to 140 keV, PSNR and SSIM of modelmix were comparable to or higher than those of model70 and model100. For attenuation correction, modelmix performed better than model100 at 70 keV level and model70 at 100 keV level (P<0.010) but comparably to modelMAR at both levels (P>0.050). Meanwhile, image noise in the spinal canal of three deep-MAR models at 100 keV level were lower than these of modelMAR (P<0.010). The scores of subjective image quality for modelmix were comparable to or higher than those of model70 and model100.

CONCLUSION

With better image quality across broader energy levels, multiple energy CT data are recommended to be comprised in the training of deep-MAR model for postoperative lumbar CT scanning.