Data-Driven Approaches for the Automatic Dental Pulp Exposure Detection: A Study on Periapical Oral Radiographs

Abstract

Pulp exposure is a condition in which the pulp chamber of the tooth is exposed, usually due to trauma, dental caries, or unintentional tooth fractures. If not treated right away, this exposure may lead to inflammation, infection, and eventual involvement of the apical tissues. Instead of treating the infection after it has arisen, it is usually more beneficial to stop the progression of the illness before it becomes an apical infection. To lower the risk of serious dental infections and maintain tooth vitality in this situation, early detection of pathological pulp exposure is essential. This study explores various data-driven deep learning approaches for automatic pulp exposure detection in periapical dental X-rays. We first propose a novel transfer learning-based method for accurately detecting pulp exposure in periapical X-rays. It is based on the deep learning architecture MobileNetV3-Small and obtains an accuracy value of 99.12% with a precision and recall values of 99.50% and a specificity value of 96.55%, which has surpassed various deep learning models like EfficientNet B0, B1, B2, B4, VGG16, ResNet50, ConvNext-Tiny, and DenseNet121, on a total of 1137 periapical X-rays collected across multiple dental clinics. A novel augmentation strategy was also proposed to address the problem of data imbalance in the collected data. Secondly, we perform a comparative analysis of different SSL (Self-supervised Learning) approaches with a transfer learning approach in order to understand the appropriate number of annotations and assess which techniques perform well with varying percentages of labeled data. From our experimental findings, it has been observed that with just 25%-50% of the annotated data, self-supervised approaches such as SimCLR and BYOL can perform at a high level, surpassing transfer learning approaches in performance metrics like accuracy, specificity, and precision. The code is available at https://github.com/DEVIKA1989/Automatic-Dental-Pulp-exposure-Detection.