Leveraging deep learning semantic segmentation for imaging coral skeletons

Micro-computed tomography and semantic segmentation provide insights into the structural hierarchy of biomineralized tissues, as exemplified by stony coral skeletons. Non-destructive imaging reveals mechanistic aspects of skeletal growth, including variations in porosity, density, and skeletal thickness across species and in the context of disease. Recent developments in semantic segmentation based on convolutional neural network models are poised to transform the streamlined analysis of large 3D tomography datasets, surpassing the performance of traditional segmentation methods. In this work, a series of U-Net deep learning models were trained and applied on exemplary micro-CT datasets from stony corals pertaining to Montastraea cavernosa and Porites astreoides species for the segmentation of pores and skeleton. The models were statistically evaluated, revealing that Attention U-Net was the top performer with respect to computational efficiency, accuracy, and generalizability, followed by U-Net++ and standard U-Net. Our analysis highlights accuracy limitations of U-Net-based deep learning segmentations that can result in false-positive or false-negative classifications. The segmented 3D models were utilized to perform porosity, bulk density, and thickness analyses of each dataset, revealing quantitative differences between the two species, as well as between healthy and stony coral tissue loss disease afflicted M. cavernosa coral skeletons. This work provides a framework for streamlined training and deployment of deep learning models for semantic segmentation of calcified tissues that inform our understanding of skeletogenesis and growth patterns across species and pathogenesis contexts.