Shared with The Gufo
Abstract
Objectives: To assess the accuracy of a deep learning model in the automatic detection of periapical radiolucent lesions of upper and lower jaws by comparing it with experienced radiologists’ opinion, which represents the ground truth. Material and methods: CBCT scans of 90 patients were imported into Blue sky bio software to be cropped and annotated for detection of periapical Radiolucencies. The annotated data were sent to computer science expert to use the data in training, validation and testing to evaluate the performance of a Bayesian Convolutional Neural Network (Bayesian CNN) model for the automatic detection of periapical radiolucent lesions within maxillary and mandibular cone-beam computed tomography (CBCT) images. Results: The cross-validated diagnostic performance of the Bayesian ensemble architectures. deep learning model, with ROC-AU values of 0.9839± 0.0139, The PR-AUC values further confirm strong performance, achieving (0.9933 ± 0.0052), achieved high accuracy (0.9441± 0.0283), balanced accuracy (0.9140 ± 0.0458), and F1-scores (0.9616 ± 0.0201), indicating excellent balance between sensitivity and specificity despite class imbalance. Conclusion: From a clinical perspective, these results suggest that Bayesian ResNet-18 Architecture deep learning model function as balanced diagnostic tools with high sensitivity while maintaining acceptable specificity in computer aided detection of periapical Radiolucencies on CBCT images. ResNet-18 provides also showed an optimal balance between diagnostic performance and computational efficiency, this makes it highly suitable for most clinical deployment scenarios.
