IADR Abstract Archives
Cross-center Generalizability of Deep Neural Networks for Tooth Classification
Objectives: Convolutional neural networks are widely used in medical imaging and become increasingly popular in dentistry. However, so far the generalizability of these highly complex models was sparsely investigated. We evaluated the performance of a tooth detection and classification model, when trained and evaluated on data stemming from two different dental centers, one located in Germany (Charité – Universitätsmedizin Berlin) and the other located India (King George Medical University, (KGMU), Lucknow).
Methods: We applied RetinaNet, a one-stage object detection neural network, for tooth detection and classification on panoramic dental x-rays. The x-ray imagery was provided by Charité (1690 images) and KGMU (1324 images). Using an online annotation tool, dental experts drew bounding boxes around each tooth in each panoramic and assigned the tooth name (FDI nomenclature). First, we trained RetinaNet on the Charité dataset (1000 training set, 290 validation set) and predicted on two test sets, one from Charité (400 images) and one from KGMU (814 images). Then we randomly replaced 400 images in the training set and 110 images in the validation set with KGMU radiographs, hence keeping the size of the training and validation set fixed. We thereafter retrained RetinaNet and used the new model to predict on the two hold out test sets again. For each experiment we computed precision (predictive positive value) and recall (sensitivity). We used Python and the deep learning library Keras.
Results: The model trained on the Charité training set and evaluated on the Charité test set yielded a precision and recall of 0.98 and 0.98, respectively. The same model yielded a precision and recall of 0.96 and 0.94, on the KGMU test set, respectively. After including the KGMU data into the training process precision and recall increased for the KGMU test set (0.97 and 0.97), and decreased for the Charité test set (0.97 and 0.97).
Conclusions: The generalizability of convolutional neural networks is affected by the manifold of the input data they are trained with. For tooth detection, though, and within the assessed datasets, generalizability was relatively high.
Methods: We applied RetinaNet, a one-stage object detection neural network, for tooth detection and classification on panoramic dental x-rays. The x-ray imagery was provided by Charité (1690 images) and KGMU (1324 images). Using an online annotation tool, dental experts drew bounding boxes around each tooth in each panoramic and assigned the tooth name (FDI nomenclature). First, we trained RetinaNet on the Charité dataset (1000 training set, 290 validation set) and predicted on two test sets, one from Charité (400 images) and one from KGMU (814 images). Then we randomly replaced 400 images in the training set and 110 images in the validation set with KGMU radiographs, hence keeping the size of the training and validation set fixed. We thereafter retrained RetinaNet and used the new model to predict on the two hold out test sets again. For each experiment we computed precision (predictive positive value) and recall (sensitivity). We used Python and the deep learning library Keras.
Results: The model trained on the Charité training set and evaluated on the Charité test set yielded a precision and recall of 0.98 and 0.98, respectively. The same model yielded a precision and recall of 0.96 and 0.94, on the KGMU test set, respectively. After including the KGMU data into the training process precision and recall increased for the KGMU test set (0.97 and 0.97), and decreased for the Charité test set (0.97 and 0.97).
Conclusions: The generalizability of convolutional neural networks is affected by the manifold of the input data they are trained with. For tooth detection, though, and within the assessed datasets, generalizability was relatively high.