P-65934199-236 - The segmentation of the mandible with convolutional neuronal networks for preoperative 3D-planning

Abstract

Objectives Multiple sophisticated automatic and semi-automatic techniques for human bone structure segmentation in CT scans have been developed in order to simplify and speed up diagnostic tasks, for example, preoperative planning and 3D-printing. We created a segmentation method based on convolutional neuronal networks (CNN) for surgical mandible 3D planning and 3D-Printing. In this study we investigated the accuracy of a neuronal network bases segmentation method. Methods Facial CTs at the Inselspital, Bern University Hospital (Switzerland) were used. The segmentation of 106 mandibles was performed manually. These 106 datasets were then divided into a training set, consisting of 96 CT scans, and a validation set with 10 CT scans. Three different, modified CNN were trained to recognize and segment the mandible using the training set. The trained networks then analyzed the test data set. These Results were then compared with the manually segmented data. Results The neural network was able to identify the mandible in all 10 different test data sets. We found an average dice coefficient of 0.961 with an accuracy of 0.999. Conclusion The results show that a trained CNN has developed an amazing ability to recognize anatomical structures such as the jaw in a CT data set. The upper and lower jaw are separated automatically. We observed that artefacts were filtered out by the neural network. This can and will greatly simplify the preparation of data sets for 3D planning and 3D printing. Multiple sophisticated automatic and semi-automatic techniques for human bone structure segmentation in CT scans have been developed in order to simplify and speed up diagnostic tasks, for example, preoperative planning and 3D-printing. We created a segmentation method based on convolutional neuronal networks (CNN) for surgical mandible 3D planning and 3D-Printing. In this study we investigated the accuracy of a neuronal network bases segmentation method. Facial CTs at the Inselspital, Bern University Hospital (Switzerland) were used. The segmentation of 106 mandibles was performed manually. These 106 datasets were then divided into a training set, consisting of 96 CT scans, and a validation set with 10 CT scans. Three different, modified CNN were trained to recognize and segment the mandible using the training set. The trained networks then analyzed the test data set. These Results were then compared with the manually segmented data. The neural network was able to identify the mandible in all 10 different test data sets. We found an average dice coefficient of 0.961 with an accuracy of 0.999. The results show that a trained CNN has developed an amazing ability to recognize anatomical structures such as the jaw in a CT data set. The upper and lower jaw are separated automatically. We observed that artefacts were filtered out by the neural network. This can and will greatly simplify the preparation of data sets for 3D planning and 3D printing.