Development and evaluation of a deep learning model to reduce exomass-related metal artefacts in cone-beam CT: an ex vivo study using porcine mandibles

Abstract

Objectives: To develop and evaluate a deep learning (DL) model to reduce metal artefacts originating from the exomass in cone-beam CT (CBCT) of the jaws.

Methods: Five porcine mandibles, each featuring six tubes filled with a radiopaque solution, were scanned using four CBCT units before and after the incremental insertion of up to three titanium, titanium-zirconium, and zirconia dental implants in the exomass of a small field of view. A conditional denoising diffusion probabilistic model, using DL techniques, was employed to correct axial images from exomass-related metal artefacts across the CBCT units and implant scenarios. Three examiners independently scored the image quality of all datasets, including those without an implant (ground truth), with implants in the exomass (original), and DL-generated ones. Quantitative analysis compared contrast-to-noise ratio (CNR) to validate artefact reduction using repeated measures analysis of variance in a factorial design followed by Tukey test (α = .05).

Results: The visualisation of the hard tissues and overall image quality was reduced in the original and increased in the DL-generated images. The score variation observed in the original images was not observed in the DL-generated images, which generally scored higher than the original images. DL-generated images revealed significantly greater CNR than both the ground truth and their corresponding original images, regardless of the material and quantity of dental implants and the CBCT unit (P < .05). Original images revealed significantly lower CNR than the ground truth (P < .05).

Conclusions: The developed DL model using porcine mandibles demonstrated promising performance in correcting exomass-related metal artefacts in CBCT, serving as a proof-of-principle for future applications of this approach.

Keywords: artefacts; artificial intelligence; cone-beam CT; deep learning; dental implants.

Figure 1.

Ground truth setup (no dental implant in the exomass). (A) Lateral and top photographs of the porcine mandible: the superimposed empty shapes represent the small field of view. The resinous plugs, used to prevent clogging of the implant perforation, were removed during the CBCT scanning. Representative CBCT axial reconstructions from four different units: (B) 3D Accuitomo 170, (C) Axeos, (D) ProMax 3D Max, (E) X800.

Figure 2.

Representative original and deep learning (DL)-corrected CBCT reconstructions at the same axial level as a function of the material and quantity of dental implants in the exomass from the 3D Accuitomo 170 CBCT unit.

Figure 3.

Representative original and deep learning (DL)-corrected CBCT reconstructions at the same axial level as a function of the material and quantity of dental implants in the exomass from the Axeos CBCT unit.

Figure 4.

Representative original and deep learning (DL)-corrected CBCT reconstructions at the same axial level as a function of the material and quantity of dental implants in the exomass from the ProMax 3D Max CBCT unit.

Figure 5.

Representative original and deep learning (DL)-corrected CBCT reconstructions at the same axial level as a function of the material and quantity of dental implants in the exomass from the X800 CBCT unit.

Figure 6.

Radar charts of the average scores for hard tissue visualisation and overall image quality for both original and deep learning (DL)-corrected images in the presence of dental implants made from three materials and in three quantities in the exomass for the four CBCT units, compared to the ground truth (no dental implant in the exomass). Scores higher than the ground truth (outer from score 3) indicate improved conditions, while scores lower than the ground truth (inner from score 3) indicate worsened conditions.