IADR Abstract Archives
3D Automated Segmentation of Three Craniofacial Structures Using Machine Learning
Objectives: Accurate segmentation of craniofacial structures is crucial for various clinical and research applications, particularly in orthodontics and maxillofacial surgery. Automated techniques, particularly convolutional neural networks (CNNs), have emerged as an efficient and cost-effective means of generating accurate segmentations from CBCT data. This study presents a segmentation analysis of the maxilla, mandible, and maxillary sinus using a CNN architecture, specifically the 3D U-Net, implemented within the MIScnn framework.
Methods: We utilized 355 cone-beam computed tomography (CBCT) images initially segmented using Relu® Creator and further refined manually to obtain ground truth data. Comprehensive data preprocessing, augmentation, and the CNN model was trained over 385 epochs, with early stopping implemented once the loss values plateaued to prevent overfitting.
Results: The performance of the segmentation model was evaluated using the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics. For the maxilla, the DSC averaged 0.9143 (maximum: 0.9382, minimum: 0.8742, SD: 0.0159), while the IoU averaged 0.8426 (maximum: 0.8836, minimum: 0.7765, SD: 0.0267). The mandible showed the highest accuracy, with a DSC average of 0.9699 (maximum: 0.9752, minimum: 0.9629, SD: 0.0033) and an IoU average of 0.9415 (maximum: 0.9515, minimum: 0.9284, SD: 0.0063). The maxillary sinus exhibited a DSC average of 0.9598 (maximum: 0.9719, minimum: 0.9336, SD: 0.0091), and an IoU average of 0.9228 (maximum: 0.9453, minimum: 0.8754, SD: 0.0166).
Conclusions: These results indicate a high level of accuracy for segmenting all three anatomical structures, with particularly robust performance for the mandible, evidenced by its highest DSC and lowest standard deviation across both DSC and IoU metrics. Early stopping allowed efficient training without overfitting, achieving reliable segmentation with minimal loss degradation. Future work will further optimize the model architecture and expand the dataset for broader clinical applicability.
Methods: We utilized 355 cone-beam computed tomography (CBCT) images initially segmented using Relu® Creator and further refined manually to obtain ground truth data. Comprehensive data preprocessing, augmentation, and the CNN model was trained over 385 epochs, with early stopping implemented once the loss values plateaued to prevent overfitting.
Results: The performance of the segmentation model was evaluated using the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics. For the maxilla, the DSC averaged 0.9143 (maximum: 0.9382, minimum: 0.8742, SD: 0.0159), while the IoU averaged 0.8426 (maximum: 0.8836, minimum: 0.7765, SD: 0.0267). The mandible showed the highest accuracy, with a DSC average of 0.9699 (maximum: 0.9752, minimum: 0.9629, SD: 0.0033) and an IoU average of 0.9415 (maximum: 0.9515, minimum: 0.9284, SD: 0.0063). The maxillary sinus exhibited a DSC average of 0.9598 (maximum: 0.9719, minimum: 0.9336, SD: 0.0091), and an IoU average of 0.9228 (maximum: 0.9453, minimum: 0.8754, SD: 0.0166).
Conclusions: These results indicate a high level of accuracy for segmenting all three anatomical structures, with particularly robust performance for the mandible, evidenced by its highest DSC and lowest standard deviation across both DSC and IoU metrics. Early stopping allowed efficient training without overfitting, achieving reliable segmentation with minimal loss degradation. Future work will further optimize the model architecture and expand the dataset for broader clinical applicability.