A deep learning system known as the Mask RCNN–Self Attention framework has demonstrated preliminary feasibility for automated cone-beam computed tomography segmentation and three-dimensional volumetric assessment of lateral incisor root resorption associated with impacted maxillary canines. The experimental artificial intelligence framework was engineered to produce three-dimensional masks capable of delineating complex dental and bony structures, allowing researchers to preliminarily quantify the extent of dental root-volume loss resulting from canine impactions on three-dimensional dental scans.
The primary objective of the investigation was to evaluate the technical feasibility and accuracy of using the Mask RCNN–Self Attention architecture for automated dental image segmentation and root evaluation. Both the artificial intelligence segmentation masks and the resulting volumetric measurements were directly compared against these manual annotations to systematically assess segmentation performance, volumetric differences, exploratory classification, and regression associations.
The experimental cohort consisted of cone-beam computed tomography scans collected from forty-two patients presenting with impacted maxillary canines. For neural network development, the patient imaging data were randomly partitioned into training, validation, and test subsets according to a proportional seven-to-two-to-one distribution. To provide an anatomical baseline for comparison, the researchers also retrospectively analyzed cone-beam computed tomography scans from ten bilaterally unaffected control individuals, enabling structural evaluations between impacted and unaffected anatomy.
Model training was conducted using cross-sectional image slices to guide the Mask RCNN–Self Attention network in segmenting four designated anatomical structures: the maxilla, the impacted maxillary canine, the adjacent lateral incisor, and the first premolar. The automated network demonstrated high segmentation accuracy, achieving a Dice similarity coefficient of 93.60 percent for the maxilla and 92.50 percent across individual teeth, confirming robust automated delineation. In boundary delineation assessments, the model recorded a 95th percentile Hausdorff distance equal to or less than 2.0 millimeters. In addition to high geometric precision, the automated segmentation procedure was faster than manual segmentation.
To quantify the extent of tissue destruction, the study implemented a voxel-based volumetric root-volume loss grading scheme that categorized lateral incisor root resorption severity into five stages designated from V0 to V4. Among the analyzed patient cohort, lateral incisor root-volume loss was classified as V0 in three cases (7.1 percent), V1 in sixteen cases (38.1 percent), V2 in fourteen cases (33.3 percent), V3 in eight cases (19.0 percent), and V4 in one single case (2.4 percent). Although advanced root loss representing categories V3 and V4 was observed, the vast majority of examined cases clustered within the mild to moderate V1 and V2 stages.
Volumetric comparisons revealed that both the maxilla and lateral incisor roots exhibited significantly smaller volumes on the impacted side compared to contralateral control structures, a difference that reached statistical significance with a P value of less than 0.01.
In exploratory statistical association analyses, impacted side status and patient sex were significantly associated with lateral incisor root volume, whereas overall maxillary volume provided no additional explanatory power within the regression models. Furthermore, the researchers evaluated an exploratory support vector machine classifier designed to identify resorption status; while this exploratory machine learning model achieved an area under the curve metric, it demonstrated limited sensitivity for identifying root resorption-positive cases. Overall, the authors concluded that the proposed framework provides preliminary feasibility for automated volumetric root resorption assessment.
Subject of Research: Automated CBCT segmentation and volumetric assessment of root resorption
Article Title: Automated segmentation and volumetric assessment for root resorption of lateral incisor associated with impacted canine using Mask RCNN-SA framework
Article References: Guo, C., Qin, H., Yang, F., Zhang, Y., Zhang, Y., Li, R., Qiu, X., & Shen, Y. (2026). Automated segmentation and volumetric assessment for root resorption of lateral incisor associated with impacted canine using Mask RCNN-SA framework. BMC Oral Health. https://doi.org/10.1186/s12903-026-10304-7
Image Credits: AI Generated
DOI: 10.1186/s12903-026-10304-7
Keywords: Mask RCNN-SA, CBCT segmentation, root resorption, lateral incisors, impacted maxillary canines, volumetric assessment, voxel-based grading, deep learning, maxilla segmentation, first premolar, Dice similarity coefficient, Hausdorff distance, support vector machine, three-dimensional masks, computed tomography, manual segmentation, anatomical volume
News Source: Blake Davidson. (November 19, 2026). Automated Deep Learning Framework Assesses Lateral Incisor Root Resorption on CBCT. Scienmag.



