Convolutional neural networks may be effective at detecting fused-rooted mandibular second molars on two-dimensional X-rays prior to root canal treatment.

Researcher found the ResNet18 model in combination with multiangle projection had the highest overall accuracy (79.25%) compared with other models and was able to effectively identify apical separation, asymmetrical cervical isthmus, apical isthmus, and apical merge. The model’s accuracies were 81.13%, 86.79%, and 90.57% in specific classifications, respectively. Additionally, convolutional neural networks outperformed the root canal morphology classification accuracy of endodontic residents—who had a 60.38% average accuracy (P < .05).

Researchers used micro-computed tomography reconstruction images and two-dimensional X-ray projection images of mandibular second molars. The molars’ ground truth classifications—including 109 merging, 119 symmetrical, and 43 asymmetrical molar types—were determined by the micro-CT reconstruction images, according to a study published in the Journal of Endodontics.

Traditional augmentation techniques were used to amplify the X-ray images, and several pretrained models (VGG19, ResNet50, and EfficientNet-b5) were employed to identify the molars’ root canal morphologies. The researchers then compared the classification of the molars performed by endodontic residents with those performed by the convolutional neural networks.

The researchers concluded that convolutional neural networks may have the potential to aid clinicians in the diagnosis of fused-rooted mandibular second molars.

A full list of disclosures can be found in the original study.