Machine learning framework for maxillofacial preoperative planning

Abstract: Artificial intelligence in biomedical image processing is approaching human performance at object localization while saving immense amounts of time for the physicians. These AI algorithms have the potential to automatically segment anatomical structures for preoperative planning. However, there are currently no tools such tools on the market. This study propose a framework of generating effective machine learning algorithms, applicable on different anatomical structures, to be used to increase automation in virtual surgical planning software. In this study a limited data set consisting of 34 CT image volumes was used to generate labelled training data to a Convolutional Neural Network (CNN) called Unet. The networks were evaluated with metric evaluation as well as visually evaluated. The framework produced two networks for automatic segmentation, one for the orbital bone and one for the mandibular bone. The orbital automation made useful segmentations ready for 3D printing while the mandible automation needs more work to be able to make printable segmentations. In conclusion this framework provides a viable approach of generating anatomical models for virtual surgical planning.