Objective: To apply the technique of deep learning on a small dataset of panoramic images for the detection and segmentation of the mental foramen (MF). Study design: In this study we used in-house dataset created within the School of Dental Medicine, Tel Aviv University. The dataset contained randomly chosen and anonymized 112 digital panoramic X-ray images and corresponding segmentations of MF. In order to solve the task of segmentation of the MF we used a single fully convolution neural network, that was based on U-net as well as a cascade architecture. 70% of the data were randomly chosen for training, 15% for validation and accuracy was tested on 15%. The model was trained using NVIDIA GeForce GTX 1080 GPU. The SPSS software, version 17.0 (Chicago, IL, USA) was used for the statistical analysis. The study was approved by the ethical committee of Tel Aviv University. Results: The best results of the dice similarity coefficient ( DSC), precision, recall, MF-wise true positive rate (MFTPR) and MF-wise false positive rate (MFFPR) in single networks were 49.51%, 71.13%, 68.24%, 87.81% and 14.08%, respectively. The cascade of networks has shown better results than simple networks in recall and MFTPR, which were 88.83%, 93.75%, respectively, while DSC and precision achieved the lowest values, 31.77% and 23.92%, respectively. Conclusions: Currently, the U-net, one of the most used neural network architectures for biomedical application, was effectively used in this study. Methods based on deep learning are extremely important for automatic detection and segmentation in radiology and require further development.
Objective: To apply the technique of deep learning on a small dataset of panoramic images for the detection and segmentation of the mental foramen (MF). Study design: In this study we used in-house dataset created within the School of Dental Medicine, Tel Aviv University. The dataset contained randomly chosen and anonymized 112 digital panoramic X-ray images and corresponding segmentations of MF. In order to solve the task of segmentation of the MF we used a single fully convolution neural network, that was based on U-net as well as a cascade architecture. 70% of the data were randomly chosen for training, 15% for validation and accuracy was tested on 15%. The model was trained using NVIDIA GeForce GTX 1080 GPU. The SPSS software, version 17.0 (Chicago, IL, USA) was used for the statistical analysis. The study was approved by the ethical committee of Tel Aviv University. Results: The best results of the dice similarity coefficient ( DSC), precision, recall, MF-wise true positive rate (MFTPR) and MF-wise false positive rate (MFFPR) in single networks were 49.51%, 71.13%, 68.24%, 87.81% and 14.08%, respectively. The cascade of networks has shown better results than simple networks in recall and MFTPR, which were 88.83%, 93.75%, respectively, while DSC and precision achieved the lowest values, 31.77% and 23.92%, respectively. Conclusions: Currently, the U-net, one of the most used neural network architectures for biomedical application, was effectively used in this study. Methods based on deep learning are extremely important for automatic detection and segmentation in radiology and require further development.