A reliable deep-learning-based method for alveolar bone quantification using a murine model of periodontitis and micro-computed tomography imaging

Abstract

Objectives: This study focuses on artificial intelligence (AI)-assisted analysis of alveolar bone for periodontitis in a mouse model with the aim to create an automatic deep-learning segmentation model that enables researchers to easily examine alveolar bone from micro-computed tomography (µCT) data without needing prior machine learning knowledge.

Methods: Ligature-induced experimental periodontitis was produced by placing a small-diameter silk sling ligature around the left maxillary second molar. At 4, 7, 9, or 14 days, the maxillary bone was harvested and processed with a µCT scanner (µCT-45, Scanco). Using Dragonfly (v2021.3), we developed a 3D deep learning model based on the U-Net AI deep learning engine for segmenting materials in complex images to measure alveolar bone volume (BV) and bone mineral density (BMD) while excluding the teeth from the measurements.

Results: This model generates 3D segmentation output for a selected region of interest with over 98 % accuracy on different formats of µCT data. BV on the ligature side gradually decreased from 0.87 mm³ to 0.50 mm³ on day 9 and then increased to 0.63 mm³ on day 14. The ligature side lost 4.6 % of BMD on day 4, 9.6 % on day 7, 17.7 % on day 9, and 21.1 % on day 14.

Conclusions: This study developed an AI model that can be downloaded and easily applied, allowing researchers to assess metrics including BV, BMD, and trabecular bone thickness, while excluding teeth from the measurements of mouse alveolar bone.

Clinical significance: This work offers an innovative, user-friendly automatic segmentation model that is fast, accurate, and reliable, demonstrating new potential uses of artificial intelligence (AI) in dentistry with great potential in diagnosing, treating, and prognosis of oral diseases.

Keywords: Deep learning; Micro-CT; Periodontitis.

Figure 1.

Alignment and density histogram of data for deep learning model. (A-B) An example illustrating the effect of 3D rotation on ABL measurement. (A) The five colored lines correspond to the ABL of the distal part of the first molar, the mesial and distal part of the second molar, and the mesial part of the third molar. (B) After a slight rotation, the start and end points of the five colored lines do not accurately measure the CEJ and the most coronal end of the crestal bone. (C) Standard positions: blue: coronal plane is symmetrical; pink: sagittal plane parallels maxillary second molar axis; green: transverse plane parallels maxillary bone. (D) Density histogram of three file formats. x-axis: density value; y-axis: frequency.

Figure 2.

U-Net based automatic segmentation model training on DICOM format (A) Architecture and training of the deep-learning U-Net model. After making trial runs with different parameter combinations, the network parameters values were selected for the quality of inference on unseen data (>98% accuracy) and reasonable training time. (B) U-Net model training log. (C-D) Images at the 3D level and cross-section images of the axial palate and coronal planes of the input data (C) and the corresponding output segmentation result (D). Green: background; orange: bone; blue: teeth.

Figure 3.

Selection of ROIs and application of automatic segmentation model. (A) ROIs. Two spherical areas with a 2-mm diameter are centered on the root bifurcation of the second molar to include the area of the alveolar bone affected by ligature-induced periodontitis: the entire maxillary second molar, the distal root of the maxillary first molar, the mesial root of the maxillary third molar, and the corresponding alveolar bone. (B-C) The transverse and coronal plane of segmentation results, respectively. Left: control side; right: ligature side. Green: background; orange: bone; blue: teeth. (D-E) 3D auto-segmentation results. (F) 3D auto-segmentation results of bone only. Left: control side; right: ligature side.

Figure 4.

Alveolar bone loss in the ligature-induced experimental periodontitis model. (A) Schematic experimental plan for ligature-induced periodontitis. (B) Representative images of periodontal bone loss at days 4, 7, 9, and 14. Bone density values are represented using a color scale —blue/violet: lowest density; red, highest intensity. (C) Alveolar bone volume (mm³) in the selected spherical ROI, excluding the teeth and background. We did not observe statistical differences in bone volume from the four groups on the control side (left). The alveolar bone volume of the ligature side was the lowest on days 7 and 9 and had increased on day 14. (D) BV ratio (ligature side/control side). The BV ratio continued to decrease on days 7 and 9 and increased on day 14 but remained below levels on day 4. A pairwise comparison using a one-way ANOVA was performed for statistical significance. *P < 0.05; **P < 0.01; NS, not significant. Data (mean ± SD) are biological replicates (n=3).

Figure 5.

Bone analysis of ligature-induced periodontitis after automatic segmentation. All measurements collected from the 2 mm spherical ROIs exclude the teeth and background. (A) BMD of periodontal bone at day 4, 7, 9, and 14. (B) BMD ratio (ligature side/control side). BMD ratio decreased starting on day 4 and continued to decrease on days 9 and 14, although no significant difference was observed between days 7 and 9. (C) Alveolar bone surface area (mm²) showed no statistical difference among the four groups on the control side (left). On the ligature side (right), the alveolar bone surface was smallest on days 7 and 9 and increased on day 14 but remained below day 4 levels. (D) Alveolar bone surface area ratio (ligature side/control side). (E) Average Tb/Th (mm) shows no statistical difference among the four groups on the control side (left). On the ligature side (right), average trabecular thickness decreased from day 4 to 7, with no significant changes among days 7, 9, and 14. (F) Average Tb/Th (ligature side/control side). A pairwise comparison using a one-way ANOVA was performed for statistical significance. *P < 0.05; **P < 0.01, ***P < 0.001; NS, not significant. Data (mean ± SD) are from biological replicates (n=3).