Towards the analysis of coral skeletal density-banding using deep learning

Abstract

Abstract: X-ray micro–computed tomography (µCT) is increasingly used to record the skeletal growth banding of corals. However, the wealth of data generated is time consuming to analyse for growth rates and colony age. Here we test an artificial intelligence (AI) approach to assist the expert identification of annual density boundaries in small colonies of massive Porites spanning decades. A convolutional neural network (CNN) was trained with µCT images combined with manually labelled ground truths to learn banding-related features. The CNN successfully predicted the position of density boundaries in independent images not used in training. Linear extension rates derived from CNN-based outputs and the traditional method were consistent. In the future, well-resolved 2D density boundaries from AI can be used to reconstruct density surfaces and enable studies focused on variations in rugosity and growth gradients across colony 3D space. We recommend the development of a community platform to share annotated images for AI.