CT Segmentation of Dinosaur Fossils by Deep Learning

Abstract

Recently, deep learning has reached significant advancements in various image-related tasks, particularly in medical sciences. Deep neural networks have been used to facilitate diagnosing medical images generated from various observation techniques including CT (computed tomography) scans. As a non-destructive 3D imaging technique, CT scan has also been widely used in paleontological research, which provides the solid foundation for taxon identification, comparative anatomy, functional morphology, etc. However, the labeling and segmentation of CT images are often laborious, prone to error, and subject to researchers own judgements. It is essential to set a benchmark in CT imaging processing of fossils and reduce the time cost from manual processing. Since fossils from the same localities usually share similar sedimentary environments, we constructed a dataset comprising CT slices of protoceratopsian dinosaurs from the Gobi Desert, Mongolia. Here we tested the fossil segmentation performances of U-net, a classic deep neural network for image segmentation, and constructed a modified DeepLab v3+ network, which included MobileNet v1 as feature extractor and practiced an atrous convolutional method that can capture features from various scales. The results show that deep neural network can efficiently segment protoceratopsian dinosaur fossils, which can save significant time from current manual segmentation. But further test on a dataset generated by other vertebrate fossils, even from similar localities, is largely limited.