Geometric Fidelity of Magnetic Resonance Imaging and Computed Tomography-Derived Virtual 3D Models of Porcine Cadaver Mandibles: Conventional Versus Artificial Intelligence-Based Segmentation

Abstract

Purpose: The workflow for virtual surgical planning (VSP) and the application of CAD/CAM (computer-aided design/computer-aided manufacturing) procedures are mainly based on computed tomography (CT) derived DICOM data sets. Alternatively, this study aims to preclinically illuminate the feasibility of a magnetic resonance imaging (MRI) based workflow and the impact of artificial intelligence (AI) based segmentation on the required fidelity on basic 3D geometry acquisition.

Materials and methods: Porcine cadaver mandibles were imaged with CT and a T1-weighted MRI sequence. The resulting DICOM data sets were segmented conventionally (Mimics Medical 17.0, Materialize; Belgium) and with AI-based segmentation software (ImFusion Labels and Suite, Version 2.19.2, ImFusion; Germany). The four standard tessellation language (STL) files were superimposed with a corresponding reference model derived from an optic scan (Artec Space Spider, Artec 3D; Luxembourg) and the following parameters were analysed: Hausdorff distance (HD), mean surface distance (MSD), root mean square distance (RMSD), time.

Results: In comparison to the reference model, all four parameters were significantly (P 0.001) better for the CT imaging and the AI-based segmentation. MRI-derived AI-based segmentation reached the fidelity of CT imaging data sets and conventional segmentation (HD, MSD, and RMSD each P >0.05).

Conclusion: The use of AI-based segmentation software proved to be useful and feasible for MRI-derived data sets, and generated the desired 3D geometry more quickly while maintaining the necessary quality. Nevertheless, the results for the CT were still better and remain yet the standard.

Keywords: artificial intelligence; computed tomography; computer-assisted surgery; image processing; magnetic resonance imaging; segmentation.

Fig 1

Study design.

Fig 2a and b

Porcine cadaver mandible after (a) ‘dynamic region growing’ (axial, MRI) and (b) ‘thresholding’ (axial, CT).

Fig 3a and b

ImFusion Labels in axial layers (a) setting of the background markers and (b) setting of the foreground markers.

Fig 4

Colour map of the surface deviation measurements [mm] of the segmented porcine cadaver mandibles compared to the optical 3D scan (histogram showing the distributions of the deviations of the corresponding points in mm: red colour indicating higher deviation; blue colour indicating lower deviation; MeshLab (Version 2021.07)).

Fig 5a to d

Comparison of different parameters depending on the used modality and the method of segmentation: (a) Hausdorff distances (HD), (b) mean square distances (MSD), (c) root mean square distance (RMSD), and (d) time.

Fig 6

Bland–Altman plots of the differences between the compared combinations of applied segmentation algorithms/modalities and the semi-automatic segmentation of CT data sets with regard to mean surface distances (MSD).