Study Shows Promise in Automated Endodontic Difficulty Assessment Using Deep Learning Models

A study constructed and validated a deep learning model for automating the assessment of endodontic case complexity from periapical radiographs. The study compiled a dataset of 1,386 radiographs from two clinical sites. The study employed convolutional neural networks including VGG16 and ResNet18, training them through transfer learning from ImageNet weights and self-supervised contrastive learning on 20,295 unlabeled dental radiographs. Evaluation via 10-fold cross-validation showed promising results, with the VGG16 model achieving 87.62% accuracy in classifying difficulty. Regression predicted scores with a minimal error of ±3.21. Notably, all models surpassed human examiners in performance, revealing potential in automated endodontic difficulty assessment.