Radiographic Archetypes of Periapical Lesions from Deep Embeddings: A Baseline Weighted ResNet-18 Explainability Study on 3,924 Annotated Periapical Images.
Authors
Affiliations (1)
Affiliations (1)
- Department of Restorative Dental Science, Division of Endodontics, College of Dentistry, King Saud University, Riyadh, Saudi Arabia, Phone: +966 114677420, e-mail: [email protected].
Abstract
Here, we present a simple endodontic imaging procedure that leverages expert extensible markup language (XML) annotations of publicly available periapical radiographs. To develop a machine-learning model for automated severity classification, this work used a publicly accessible dataset of periapical radiographs with labeled lesions. After building a lesion-level index from XML annotations and aggregating bounding boxes into image-level features, 3,924 radiographs were split into stratified training, validation, and test sets. A weighted residual network (ResNet)-18 model was fine-tuned with a class-balanced loss, achieving optimal performance at the best validation checkpoint. To identify radiographic trends and sources of classification error, model predictions were further examined using confusion metrics, clustering of penultimate embeddings, and visual explanations. Lesion-level analysis identified 6,741 annotated locations across 3,924 radiographs, showing frequent multi-lesion presentations and a predominance of moderate cases. A weighted ResNet-18 model demonstrated middling classification performance, with strong recall for mild lesions but much decreased detection capability for intermediate and severe disease. Interpretability evaluations revealed that the model concentrated on lesion locations for accurate predictions, whereas embedding visualizations and clustering demonstrated that lesion severity resides on a continuum rather than in distinct categories. We constructed a weighted ResNet-18 model to identify periapical lesion (PL) severity in three categories using the Mendeley periapical radiograph dataset. Our results imply that the model's deep feature representations exhibit distinct radiographic embedding patterns and a well-organized error structure. The application of convolutional neural networks (CNNs) to two-dimensional periapical and panoramic radiographs (PRs) for diagnosing "Periapical Lesions" with the complexity of anatomical structures and overlapping artifacts in the oral and maxillofacial (OMF) region would be useful for clinicians in diagnosing and predicting oral disease when integrated with routine electronic health records (EHRs).