Intelligent Objective Osteon Segmentation Based on Deep Learning

Abstract

Histology is key to understand physiology, development, growth and even reproduction of extinct animals. However, the identification and interpretation of certain structures, such as osteons, medullary bone (MB), and Lines of Arrested Growth (LAGs), are not only based on personal judgments, but also require considerable labor for subsequent analysis. Due to the dearth of available specimens, only a few quantitative histological studies have been proceeded for limited dinosaur taxa, most of which focus primarily on their growth, namely, LAGs and other growth lines without much attention to other histological structures. Here we develop a deep convolutional neural network-based method for automated osteohistological segmentation. Raw images are firstly divided into sub-images and the borders are expanded to guarantee the osteon regions integrity. ResNet-50 is employed as feature extractor and atrous spatial pyramid pooling (ASPP) is used to capture multi-scale information. A dual-resolution segmentation strategy is designed to observe the primary and secondary osteon regions from the matrix background. Finally, a segmented map with different osteon regions is obtained. This deep convolutional neural network-based model is tested on a histological dataset derived from various taxa in Alvarezsauria, a highly specialized group of non-avian theropod dinosaurs. The results show that large-scale quantitative histological analysis can be achieved by neural network-based methods, and previously hidden information by traditional methods can be revealed. Phylogenetic mapping of osteon segmentation results suggests a developmental pathway towards miniaturized body sizes in the evolution of Alvarezsauria, which may resemble the transition from non-avian dinosaurs to birds.