IADR Abstract Archives
Implementation of Deep Learning Using a Convolutional Neural Network in Identifying Periapical Lesions
Objectives: Differentiation between granulomas and radicular cysts is important subject to decide on the treatment. Surgical biopsy and subsequent histopathological evaluation remain the gold standard for confirming the diagnosis of these periapical lesions. We created a convolutional neural network (CNN) for the detection of the periapical lesions based on digital radiograhy.
Methods: Three hundred digital radiographic images data of granulomas and radicular cysts were retrospectively gathered from a dental hospital medical records system with the histopathological confirmation, whose patient information could not be identified, and preprocessed for data normalization. Transfer learning used CNN architectures BaseNet, Inception V4 and NASNet for the larger sample dataset and was refined with our secondary training dataset comprising 200 images We applied data augmentation to overcome the problem of limited patient data. This experiment designed to categorize radiographic digital images into granulomas and radicular cysts lesions.
Results: The precision, recall, F-1 score and accuracy values are 83.5%, 83.5%, 83.5%, 83.5% for BaseNet, 96.0%, 96.0%, 96.0%, 96.0% for Inception V4 and 90.5%, 90.5%, 90.5%, 90.5% respectively for Nasnet CNN architecture.
Conclusions: These results demonstrate that CNN may aid in detection for granulomas and radicular cyst. The initial results are very encouraging and the application is the first of its kind and, with further refinement, has excellent potential to be of benefit in periapical lesions differentiation.
Methods: Three hundred digital radiographic images data of granulomas and radicular cysts were retrospectively gathered from a dental hospital medical records system with the histopathological confirmation, whose patient information could not be identified, and preprocessed for data normalization. Transfer learning used CNN architectures BaseNet, Inception V4 and NASNet for the larger sample dataset and was refined with our secondary training dataset comprising 200 images We applied data augmentation to overcome the problem of limited patient data. This experiment designed to categorize radiographic digital images into granulomas and radicular cysts lesions.
Results: The precision, recall, F-1 score and accuracy values are 83.5%, 83.5%, 83.5%, 83.5% for BaseNet, 96.0%, 96.0%, 96.0%, 96.0% for Inception V4 and 90.5%, 90.5%, 90.5%, 90.5% respectively for Nasnet CNN architecture.
Conclusions: These results demonstrate that CNN may aid in detection for granulomas and radicular cyst. The initial results are very encouraging and the application is the first of its kind and, with further refinement, has excellent potential to be of benefit in periapical lesions differentiation.