Convolutional neural network—assisted detection of pantomographic technique errors

Objective

Technique errors on dental pantomographic (panoramic) images are relatively common and can interfere with the diagnostic evaluation of the image.

Convolutional neural networks (CNNs) hold promise as a supplemental aid for automated detection and classification of image features. We hypothesize that CNNs are capable of detecting and classifying technique errors in panoramic images with clinically relevant accuracy. In this study, we aim to compare the capability of 2 CNNs and 7 optimizers to accurately recognize technique errors in panoramic images. To the best of our knowledge, this is the first study to explore the use of CNNs in the detection of technique errors on the basis of panoramic images features.

Study Design

Panoramic images were obtained from Stony Brook School of Dental Medicine's PACS, anonymized and classified manually for the presence of multiple classes of technique errors. In phase 1 of the study, we selected images that illustrate one category of error, the presence of palatoglossal air space, and implemented image augmentation for a more robust data set. Images were presented to modified VGG16 (Oxford's Visual Geometry Group 16 layer) and VGG19 CNN architectures. After each CNN was trained with an image subset, a separate validation set of images was presented to it. For each image in the validation set, accuracy of the CNN was measured by correct classification of presence or absence of technical error. Seven optimizers were compared for their accuracy in classification. In phase 2, the aim of the study will be to train a CNN to detect multiple panoramic technique errors.

Results

Current results are 86.5% accuracy using VGG16 with Adam optimizer. Final results are pending and will be presented in the conference poster.

Conclusion

Our results suggest the feasibility of VGG16 with Adam optimizer for the detection of single errors in dental panoramic images.