Artificial intelligent-driven decision-making for automating root fracture detection in periapical radiographs

Early and accurate detection of root fractures is crucial for treatment and prognosis, yet diagnosis using 2D periapical radiographs is subjective and dependent on operator experience. Although CBCT offers higher diagnostic accuracy, its artifacts, cost, and radiation limit routine use. Given AI’s growing role in dental diagnostics, this study investigates whether integrating AI—specifically an ensemble voting system of five pretrained CNNs (VGG16, VGG19, ResNet50, DenseNet121, DenseNet169)—can support reliable detection of root fractures on periapical radiographs and thereby reduce human error, particularly for less-experienced clinicians.

The paper reviews prior applications of AI and CNNs in dental imaging, including early validation of probabilistic neural networks for vertical root fracture detection on periapical images, and studies showing superior performance of AI on CBCT compared with 2D radiographs. Hu et al. reported ResNet50 outperforming other models on CBCT images and even surpassing experts. Other works demonstrated AI support in interpreting dental imagery, detecting oral conditions, and predicting issues. The authors note inconsistencies across models and modalities in the literature, with some studies favoring ResNet50 and others indicating varied performance depending on task (e.g., caries segmentation) and data complexity. These findings motivate comparing multiple CNN architectures and introducing a voting mechanism to improve robustness for 2D periapical fracture detection.

Design and ethics: In vitro study with IRB approval (MUST IRB 2022/0097). Specimens: 400 single-rooted extracted anterior teeth (maxillary and mandibular), including endodontically treated teeth, divided into fractured and intact groups (n=200 each). Horizontal root fractures were artificially created using a diamond cutting disc (#2, 0.15 × 22 mm) to simulate incomplete fractures. Imaging: Periapical radiographs captured using a size-2 photostimulable storage phosphor plate (VistaScan, Dürr Dental AG) and processed with DBSWIN software. Dataset: Balanced distribution of fractured/unfractured images. Train/validation-test split of 80:20. No data augmentation was used to avoid potential overfitting and artifact introduction. Preprocessing: Normalization (1/255), resizing to 224 × 224 × 3 pixels. Models: Five pretrained CNNs—VGG16, VGG19, ResNet50, DenseNet121, DenseNet169—were fine-tuned. Training configuration: 50 epochs; dropout 0.3; early stopping with patience 15; ReduceLROnPlateau after 10 epochs (factor 0.2, minimum learning rate 0.001); best weights saved; Adam optimizer (β1=0.9, β2=0.999). Visualization: Matplotlib used for plotting loss curves and ROC. Ensemble voting: A majority voting classifier (odd number of models; here n=5) aggregated individual model outputs to produce a final decision (majority defined as >50%+1 vote). Evaluation metrics: Precision (PPV), recall (sensitivity), specificity, training/validation loss, and ROC/AUC were used to assess diagnostic performance for both fractured and unfractured classes.

Per-model performance (from Table 1 unless otherwise noted): - VGG16: Training loss 0.09%, validation loss 0.18%; fractured: sensitivity 0.870, specificity 1.000, PPV 1.000; unfractured: sensitivity 1.000, specificity 0.870, PPV 0.886; ROC 0.990. Narrative also reports low losses and strong metrics without overfitting. - VGG19: Training loss 0.545%, validation loss 0.61%; fractured: sensitivity 0.447, specificity 0.745, PPV 0.636; unfractured: sensitivity 0.745, specificity 0.447, PPV 0.574; ROC 0.740. Indicators of potential overfitting. - ResNet50: Training loss 0.61%, validation loss 0.63%; fractured: sensitivity 0.426, specificity 0.851, PPV 0.740; unfractured: sensitivity 0.851, specificity 0.426, PPV 0.597; ROC 0.740. Bias toward unfractured class and high noise noted. - DenseNet121: Training loss 0.11%, validation loss 0.23%; fractured: sensitivity 0.894, specificity 0.936, PPV 0.933; unfractured: sensitivity 0.936, specificity 0.894, PPV 0.898; ROC 0.980. Balanced metrics with effective handling of overfitting/noise. - DenseNet169: Training loss 0.11%, validation loss 0.21%; fractured: sensitivity 0.872, specificity 0.936, PPV 0.932; unfractured: sensitivity 0.936, specificity 0.872, PPV 0.880; ROC 0.980. Improved generalization compared with shallower counterparts. Ensemble voting: Implemented majority voting across the five models to yield a final decision per image; illustrative examples show decisive outputs even when individual models disagreed.

Findings indicate that while individual CNNs vary in performance on 2D periapical radiographs, ensemble majority voting can provide more robust and consistent diagnostic decisions for root fracture detection. VGG16 and DenseNet variants performed best among the single models, whereas VGG19 and ResNet50 underperformed or showed biases/overfitting. Compared with literature favoring CBCT for fracture diagnosis, the present 2D approach demonstrates promising performance but acknowledges modality limitations. The ensemble strategy mirrors clinical consensus processes and mitigates overfitting and model variance, potentially supporting less-experienced dentists and improving diagnostic reliability. Differences from prior reports (e.g., superior ResNet50 on CBCT) are attributed to varying imaging modalities, datasets, target pathologies (vertical vs horizontal fractures), and architectures.

An AI-based system using five pretrained CNNs with a majority voting mechanism showed promising precision and sensitivity for detecting root fractures in periapical radiographs. The voting approach helped address discrepancies among individual models and enhanced decision-making, indicating potential for automated diagnostic support in dentistry.

The study was conducted in vitro on extracted single-rooted teeth with artificially created horizontal fractures, which may not fully represent clinical complexity. The use of 2D periapical radiographs limits generalizability compared with 3D modalities like CBCT. The dataset size (400 images) is modest and no data augmentation was applied, which may affect generalization. Results require validation in clinical settings, with larger datasets and inclusion of multirooted teeth.