
Machine Learning from X-ray Images Classifies Canal Curvature with 89.6% Accuracy
A model trained on 1,644 periapical images of mandibular premolars accurately classified 89.6% of canals by curvature degree, but authors see it only as a screening tool.
Image In, Curvature Degree Out
A retrospective study in the International Endodontic Journal developed an automated system that identifies a mandibular premolar with a single canal on a periapical image, isolates the canal, and assesses its curvature. The dataset included 1,644 images: 633 straight, 885 moderately, and 126 severely curved canals. The entire system was evaluated using tenfold cross-validation with the same data split at each step to prevent mixing of training and validation data.
For tooth detection, the RF-DETR model performed best (recall 97.5%, precision 93.3%, F1 95.4%). YOLOv11-seg was most effective in isolating the canal, with a Dice coefficient of 85.8% and IoU of 77.5%. The classifier then combined five geometric measures of canal shape with 2,048 features from a deep neural network.
The hybrid model achieved the highest overall accuracy, 89.6%. For the most critical group, severely curved canals, the model using only geometric measures was more sensitive: 80.6% compared to 69.1% for the hybrid. The strongest individual predictor of curvature severity was Luiten's angle, and all five geometric measures statistically differed among severity groups.
Limitations
The study is retrospective, focused on a single tooth type, and lacks external validation on other devices or populations. The two-dimensional image does not show buccolingual curvatures, and there were few severely curved canals (126).
For your practice
- Such a tool can flag canals worth further examination on CBCT before instrumentation, but it does not replace clinical judgment.
- In cases of suspected severe curvature, simpler geometric assessment is more accurate: it is more reliable than the "smarter" hybrid model for this specific group.
- On a standard image, buccolingual curvature is not visible; this applies to both humans and algorithms.