PaperlayerView exact revision on arXiv

Source-linked AI summary

Use of the Deep Learning Approach to Measure Alveolar Bone Level

Chun-Teh Lee, Tanjida Kabir, Jiman Nelson, Sally Sheng, Hsiu-Wan Meng, Thomas E. Van Dyke, Muhammad F. Walji, Xiaoqian Jiang, Shayan Shams

arXiv:2109.12115v1eess.IVcs.CVcs.LG

TL;DR

Clinician interpretation of radiographs can vary, motivating an objective approach to periodontal assessment. The paper develops a deep-learning model combining segmentation and image analysis to measure radiographic bone loss and assign stages. The model showed strong segmentation and staging performance, though further training and validation with more images are needed.

  • Problem

    Radiographic interpretation is subject to error and varies with clinician experience, while accurate bone-level measurement matters for periodontal diagnosis when comprehensive charting is unavailable.

    Source: Introduction · p. 3

  • Method

    The model combines segmentation of bone area, teeth, and the cementoenamel junction with image analysis to assess alveolar bone level and periodontal diagnosis.

    Source: Model development · p. 4

  • Results

    The proposed CAD achieved AUROC values of 0.98, 0.89, 0.90, and 0.90 for no bone loss and stages I, II, and III, respectively, and agreement with ground truth was κ=0.81.

    Source: Staging validation · p. 9

  • Takeaways & Limitations

    The model can provide a rapid, reliable preliminary periodontal diagnosis from radiographic findings; CAD analysis took 9 minutes versus 137 minutes for clinicians.

    Source: Discussion · p. 11

    The model can underestimate RBL at vertical defects, endoperiodontal lesions, and highly overlapped teeth, and needs training and validation on more images.

    Source: Discussion · p. 12

Abstract

from arXiv · show

Abstract: Aim: The goal was to use a Deep Convolutional Neural Network to measure the radiographic alveolar bone level to aid periodontal diagnosis. Material and methods: A Deep Learning (DL) model was developed by integrating three segmentation networks (bone area, tooth, cementoenamel junction) and image analysis to measure the radiographic bone level and assign radiographic bone loss (RBL) stages. The percentage of RBL was calculated to determine the stage of RBL for each tooth. A provisional periodontal diagnosis was assigned using the 2018 periodontitis classification. RBL percentage, staging, and presumptive diagnosis were compared to the measurements and diagnoses made by the independent examiners. Results: The average Dice Similarity Coefficient (DSC) for segmentation was over 0.91. There was no significant difference in RBL percentage measurements determined by DL and examiners (p=0.65). The Area Under the Receiver Operating Characteristics Curve of RBL stage assignment for stage I, II and III was 0.89, 0.90 and 0.90, respectively. The accuracy of the case diagnosis was 0.85. Conclusion: The proposed DL model provides reliable RBL measurements and image-based periodontal diagnosis using periapical radiographic images. However, this model has to be further optimized and validated by a larger number of images to facilitate its application.

Abstract

Radiographic bone-level assessment supports periodontal diagnosis, but interpretation can vary with clinician experience. The study presents a deep-learning model for periapical radiographs to assess bone level and assign bone-loss stages.

Clinical Relevance

  • Radiographic interpretation is subjective, and accuracy depends on clinician experience and knowledge.Source: Clinical Relevance · p. 2
  • The proposed model can reliably assess radiographic bone level and assign bone-loss stage.Source: Clinical Relevance · p. 2
  • The model can assist clinicians with diagnosis and review large numbers of intraoral images for quality control and research.Source: Clinical Relevance · p. 2

Introduction

  • Radiographic bone-level measurement is important for diagnosis, especially when comprehensive or longitudinal periodontal charting is unavailable.Source: Introduction · p. 3
  • Periapical radiographs are the study’s basis for a CAD model intended to compare automated assessment with clinicians’ assessments.Source: Introduction · p. 3

Materials and Methods

The authors developed a deep-learning computer-aided diagnosis model for alveolar bone-level assessment and periodontal diagnosis from intraoral radiographs. It combines segmentation of three anatomical targets with image analysis.

Model development

Annotation

Intraoral radiographs were annotated for bone area, teeth, and the cementoenamel junction. The authors compared U-Net variants and selected different configurations for bone/tooth versus CEJ segmentation.

Annotation

  • Clinical examiners annotated bone area, teeth, and other regions with polygons and the CEJ with a polyline; when the CEJ was obscured, the crown margin was annotated as CEJ.Source: Annotation · p. 4

Segmentation model

  • Different U-Net configurations using CNN, ResNet-34, or ResNet-50 encoders and adjusted hyperparameters were evaluated for the three segmentation tasks.Source: Segmentation model · p. 5
  • The U-Net’s contraction and expansion paths change feature-map dimensions, while cross-connections reuse contraction features during reconstruction.Source: Segmentation model · p. 5
  • U-Net with a ResNet-34 encoder performed best for bone-area and tooth segmentation, while U-Net with CNN blocks performed best for CEJ-line segmentation.Source: Segmentation model · p. 5, Table 1 · p. 5
  • Training used 100 epochs, batch size 8, binary cross-entropy loss, and the Adam optimizer.Source: Segmentation model · p. 6
  • The segmentation outputs were integrated for subsequent alveolar bone-level image analysis; average inference time was 53.4 ms per radiograph.Source: Segmentation model · p. 6

Alveolar bone level measurement

The measurement pipeline processes predicted masks, identifies tooth, bone, and CEJ intersections, and estimates root apices to calculate radiographic bone loss at mesial and distal sites. Tooth-level stage is assigned from the higher site stage.

Alveolar bone level measurement

Model training and validation

The model was trained and evaluated using annotated periapical radiographs, with separate data for development and additional evaluation. Examiner calibration and consensus rules established reference measurements and diagnoses.

Model training and validation

  • 693 images from 37 patients were split into 70% training, 10% validation, and 20% testing, with evaluation also using 644 images from 46 entirely separate cases.The separate patients were used to avoid data snooping bias.Source: Model training and validation · p. 8
  • Three calibrated examiners annotated and measured RBL; conflicts in stage assignment were resolved using measured RBL percentage to establish the final ground-truth stage.Before annotation, examiner DSC was at least 0.84, and RBL percentage measurements did not differ significantly (p>0.05).Source: Model training and validation · p. 8
  • For case-level evaluation, examiners assigned diagnoses using the same principles as the model, with majority rule resolving inconsistent examiner diagnoses.The model’s provisional diagnosis used RBL stages, tooth-level RBL percentages, and patient age.Source: Model training and validation · p. 8

Statistical analysis

The study assessed segmentation against reference masks, compared RBL measurements and stage assignments with examiner results, and estimated the sample size needed for measurement comparisons.

Statistical analysis

  • Segmentation was assessed with DSC, Jaccard Index, and pixel accuracy, while RBL stage assignment was evaluated using AUROC, sensitivity, specificity, accuracy, and Cohen’s kappa.RBL percentage and measurement time were compared with examiner results using Student’s t-test; stage validation allowed ±3% RBL percentage variation.Source: Statistical analysis · pp. 8–9
  • Jaccard Index measures similarity and diversity between model predictions and reference masks; pixel accuracy is the percentage of correctly classified images relative to references.DSC compares model output similarity against reference masks.Source: Statistical analysis · p. 8
  • 394 samples were required by power analysis for a two-sided t-test with 0.05 significance, 0.8 power, assumed SD 0.15, and a 0.03 noninferiority margin between examiner means.The null hypothesis specified that the difference between RBL percentage means was no larger than 0.03.Source: Statistical analysis · p. 9

Segmentation model validation

The model’s segmentation scores were high across bone area, tooth shape, and CEJ line, and its RBL staging showed strong discrimination and agreement with ground truth.

Segmentation model validation

  • For bone area, tooth shape, and CEJ line segmentation, DSC was 0.96, 0.95, and 0.91; JI was 0.93, 0.91, and 0.88; pixel accuracy was 0.96, 0.89, and 0.99, respectively.Source: Segmentation model validation · p. 9

Staging validation

  • AUROC was 0.98 for no bone loss and 0.89, 0.90, and 0.90 for RBL stages I, II, and III; CAD agreement with ground truth was κ=0.81 versus examiner κ values of 0.72, 0.66, and 0.52.Sensitivity, specificity, and accuracy were all over 0.8 for the different stages.Source: Staging validation · p. 9, Table 2 · p. 9

RBL percentage and level measurements

On the additional dataset, CAD RBL measurements did not differ significantly from examiners’ measurements and took less time; case diagnosis accuracy was 0.85. The authors describe the periapical-image model as a preliminary aid, not a replacement for clinical assessment.

RBL percentage and level measurements

Periodontitis diagnosis assignment

Discussion

  • The model may support preliminary diagnosis when comprehensive examination is missing and large-scale review, but accurate diagnosis still requires periodontal charting, radiographs, and patient history.The authors also describe reviewing radiographic interpretations for quality control and research purposes.Source: Discussion · pp. 11–12
  • The model combines RBL measurement with extent, stage, and grade assignment under the 2018 classification using periapical images, unlike previously described panoramic systems.The authors identify it as the only model with this combination based on periapical radiographs.Source: Discussion · p. 11
  • RBL percentage can mislead in teeth with short roots or wide supracrustal tissue attachment, so the model also reports CEJ-to-bone distance for clinician reference.Source: Discussion · p. 11
  • Stage accuracy was lower than extent and grade accuracy, and the model tends to underestimate RBL at vertical defects, endoperiodontal lesions, and highly overlapped teeth.Source: Periodontitis diagnosis assignment · p. 10, Discussion · p. 12

Encoder Decoder

The supplementary tables cover U-Net convolutional kernel sizes and the diagnoses of cases used to validate CAD accuracy.

Encoder Decoder

  • Supplementary Table S2 lists convolutional kernel sizes for the U-Net.The supplied passage does not state the kernel-size values.Source: Table S2 · p. 16
  • Supplementary Table S3 covers diagnoses of cases used to validate Computer-Aided Diagnosis accuracy.The supplied passage does not provide the table’s case-level contents.Source: Table S3 · p. 17

Related papers

Loading 2109.12115v1…