Segmenting beyond the imaging data: creation of anatomically valid edentulous mandibular geometries for surgical planning using artificial intelligence

Abstract

Background and objectives: Mandibular reconstruction following continuity resection due to tumor ablation or osteonecrosis remains a significant challenge in maxillofacial surgery. Virtual surgical planning (VSP) relies on accurate segmentation of the mandible, yet existing AI models typically include teeth, making them unsuitable for planning of autologous transplants dimensions aiming for reconstructing edentulous mandibles optimized for dental implant insertion. This study investigates the feasibility of using deep learning-based segmentation to generate anatomically valid, toothless mandibles from dentate CT scans, ensuring geometric accuracy for reconstructive planning.

Methods: A two-stage convolutional neural network (CNN) approach was employed to segment mandibles from computed tomography (CT) data. The dataset (n = 246) included dentate, partially dentate, and edentulous mandibles. Ground truth segmentations were manually modified to create Class III (moderate alveolar atrophy) and Class V (severe atrophy) models, representing different degrees of post-extraction bone resorption. The AI models were trained on the original (O), Class III (Cl. III), and Class V (Cl. V) datasets, and performance was evaluated using Dice similarity coefficients (DSC), average surface distance, and automatically detected anatomical curvatures.

Results: AI-generated segmentations demonstrated high anatomical accuracy across all models, with mean DSCs exceeding 0.94. Accuracy was highest in edentulous mandibles (DSC 0.96 ± 0.014) and slightly lower in fully dentate cases, particularly for Class V modifications (DSC 0.936 ± 0.030). The caudolateral curve remained consistent, confirming that baseline mandibular geometry was preserved despite alveolar ridge modifications.

Conclusions: This study confirms that AI-driven segmentation can generate anatomically valid edentulous mandibles from dentate CT scans with high accuracy. The innovation of the work is the precise adaptation of alveolar ridge geometry, making it a valuable tool for patient-specific virtual surgical planning in mandibular reconstruction.

Keywords: Computed tomography; Convolutional neural networks; Deep learning; Mandible; Medical image analysis; Reconstruction; Segmentation; Virtual surgical planning.

Fig. 1

3D surface comparison of predictions of the different models against the original ground truths. Note the generally valid surfaces in large regions of the mandible, e.g. at the ascending branches and the body of the mandible. The teeth are less accurately segmented in the so-called ground truths; thus, the original model shows weaker performance in this region. The other two models, Cl. III and Cl. V, show a valid geometry of the alveolar crest and a comparably accurate geometry apart from that region

Fig. 2

Comparison between predicted Cl. V and original ground truths. Geometries of the alveolar crests are always anatomically correct. Mandibular bodies and ascending branches are consistently captured with sufficient accuracy

Fig. 3

Comparison between original ground truths and predicted Cl. III. The comparison shows that for Cl. III predictions, only dental geometries were removed, leaving the continuous mandibular anatomy intact. For edentulous mandibles, their geometry was not altered significantly, as desired

Fig. 4

Visualization of Dice scores for the different groups of evaluations

Fig. 5

Illustrative chart of the two different curve metrics for the different groups

Fig. 6

Six sets of corresponding examples of ground truth segmentations. Gray: original datasets (O), blue: toothless datasets (Cl. III) and green: (Cl. V). The general shape of these mandibles remains unchanged, but the dentate part is manually removed (Cl. III) and the alveolar ridge further reduced (Cl. V), respectively

Fig. 7

Two-stage pipeline for image segmentation: The whole field of view is used for the first stage (left) and the resulting segmentation (orange) is used to define a matching bounding box (blue) to derive a region of interest for a detailed segmentation with the different models deriving different dental status, i.e. like the original data (green), a class III edentulous (yellow) and a class V edentulous version (red), respectively

Fig. 8

3D U-Net architecture (adapted from Pankert et al. [9]) used in the two-stage mandible segmentation pipeline. The same network architecture is used for both stages. In the first stage, and in each of the three single-label second-stage models (original, Class III, Class V), the network outputs a single binary channel (O_c=1). In the multi-label second-stage model, the architecture remains unchanged except for the final layer, which outputs three independent binary channels (O_c=3). All inputs are resampled to a resolution of 144 × 144 × 144 voxels with one intensity channel

Fig. 9

Visualization of automatically detected anatomical curves: dental curve (red) and caudolateral curve (green) for original data, modified Class III and modified Class V