Semantic image segmentation is a popular technique that involves assigning semantic labels to pixels of images and has been revolutionary in the fields of medical imaging and materials science Davydzenka et al., 2022, Wang et al., 2022). In materials science, the use of non-destructive X-ray imaging and image segmentation has facilitated the characterization of internal micro- and nanostructures, such as phase and grain boundaries, voids or cracks. For example, image segmentation has proven to be a useful tool for studying the interplay between nanoscale grain size and mechanical performance of metals or alloys in industrial settings (Dursun and Soutis, 2014, Hu et al., 2017). However, studying biological materials using X-ray tomography and image segmentation presents challenges due to complex and overlapping micro- and nanostructures, noisy image conditions, and low contrast (Duke et al., 2017). Despite this, complex biological materials have traditionally been evaluated by manual image segmentation (Dai et al., 2013, Darrow et al., 2015, Wang et al., 2015).

Micro-computed tomography (micro-CT) is an imaging technique that uses X-rays to resolve internal three-dimensional (3D) structures at a (sub)micron level (Vásárhelyi et al., 2020). With respect to materials characterization, micro-CT can be a powerful tool for determining qualitative and quantitative information such as porosity, pore size and shape distribution, tortuosity, thickness distribution, phase fractions, phase contiguity, and orientation (Maire & Withers, 2014). For instance, micro-CT has been used to study bone penetration into solid implants, proving to be a viable approach to evaluate the performance and incorporation of biomaterials (Cohen et al., 2017). A recent study developed bio-inspired micro air vehicles based on tomographic reconstructions of fly wings (Rubio & Chakravarty, 2016).

X-ray tomography has been extensively used to elucidate the microstructure, growth mechanism, and preferred orientation of biomineralized systems found in exo- and endoskeletons and other hard tissues of living organisms (Furat et al., 2019, Reznikov et al., 2020). An example of biomineralized structures in the marine environment are scleractinian corals, which primarily build their skeletons from a biomineral called aragonite, a highly ordered polymorph of calcium carbonate (CaCO3) (San et al., 2022). Studies on coral skeletons have provided insights into morphological changes related to environmental stressors (Thompson, 2022) and overall coral health (Lough & Cooper, 2011). Micro-CT imaging has been leveraged to study skeletal growth, including the impact of ecological and environmental factors on growth direction, bulk density, and calcification rates (Fouke et al., 2021). For example, micro-CT has been used to compare skeletal features and canal systems in various reef-building coral species, suggesting species-specific growth patterns and colony pattern regulations of dominant reef-building corals (Li et al., 2021). Similarly, fluctuations in environmental conditions result in seasonal banding patterns in massive scleractinian corals, as shown by X-ray imaging and image segmentation (Vincent & Sheldrake, 2025). Thus, micro-CT can serve as a powerful tool for systematically analyzing complex aspects of coral biomineralization that are not yet fully understood, such as the effects on environmental stressors and disease on skeletal growth mechanisms. This work attempts to leverage deep learning-based image segmentations of coral micro-CT data using an extensive model evaluation and selection framework to streamline porosity analyses of different coral species, as well as healthy and diseased corals.

Due to their high mineral content, coral skeletons are characterized by a high X-ray absorption coefficient, rendering them ideal samples for X-ray imaging. Various studies harnessed micro-CT techniques to investigate healthy coralline skeletons, particularly focusing on seasonal porosity fluctuations or growth bandings. Moreover, image segmentation studies of coral micro-CT data have investigated the effects of ocean acidification (Scucchia et al., 2025, Scucchia et al., 2022) and acclimatory abilities (Scucchia et al., 2023) with respect to skeletal thickness and porosity. However, little is known about how common coral diseases (e.g., Blackband Disease, Stony Coral Tissue Loss Disease, or Aspergillosis) may affect skeletal patterns and microstructure. Specifically, the use of deep learning assisted semantic segmentation remains underexplored in the field of biomaterials. Previous studies on the stony coral species Porites compressa tied skeletal growth anomalies in diseased corals to high porosity (Domart-Coulon et al., 2006, Andersson et al., 2020). Scucchia et al. applied the U-Net network architecture algorithm to segment rapid accretion deposits (RADs, also known as centers of calcification) and thickening deposits (TDs) in coral skeletons (Domart-Coulon et al., 2006).

While sample preparation protocols and tomography scan parameters should be optimized, some experimental settings are associated with sub-optimal data collection, such as inline quality control in industry or in-situ experiments at synchrotron beamlines. Even under the assumption of high-quality data collection, deep learning image segmentation techniques may provide unique advantages by accelerating and streamlining data analysis. In some cases, manual segmentation proves impossible since minor grey value variations cannot be discerned by eye.

In this study, modern deep learning image segmentation tools were tested and evaluated to automatically identify and delineate skeletal features. State-of-the-art image segmentation tools are crucial for accelerating materials characterization and quantitative analysis, enabling high-throughput analysis, and improving accuracy and reproducibility. In recent years, deep learning-based image segmentation algorithms have grown in popularity due to their exceptional performance compared to traditional segmentation methods (Minaee et al., 2022). Specifically, convolutional neural networks (CNNs) have been widely used for image segmentation tasks (Fukushima, 1980, Lecun et al., 1998, Waibel et al., 1989). The use of convolution layers provides the advantage of enhanced feature extraction, while nonlinear layers provide robustness through the ability to model nonlinear functions (Maire & Withers, 2014). In addition, pooling layers reduce dimensionality of input images for efficient memory usage. Up to our knowledge, only few studies have systematically compared deep-learning image segmentations of X-ray tomography reconstructions from biomineralized materials based on performance metrics and statistical evaluations.

A popular CNN-based deep learning network, U-Net, has excelled specifically for biomedical image segmentation (Ronneberger et al., 2015). Developed by Ronneberger et al, U-Net assumes an encoder-decoder structure consisting of a contracting and expansive path (Fig. 1). The contracting path assumes a typical architecture of a CNN, with repeated doublets of 3x3 convolutions, followed by a rectified linear unit (ReLU) and a 2x2 max pooling layer. The contracting path essentially consists of “downsampling” steps. The expansive path consists of an “upsampling” step followed by a 2x2 convolution and a doublet of 3x3 convolutions, each followed again by a ReLU. A final 1x1 convolution is then applied to map each component feature vector corresponding to the number of classes.

Notably, the use of skip connections in the U-Net architecture have allowed for recovery of fine-grained details even with complex backgrounds. Nonetheless, to meet the strict demands of segmenting medical images, variations of U-Net were created to address its shortcomings in this regard. U-Net++, introduced by Zhou et al, is an encoder-decoder network where the subnetworks are connected by nested, dense skip connections in order to reduce the semantic gap between feature maps of each sub-network (Zhou et al., 2018). Another variation is Attention U-Net, which adds an attention gate (AG) model to eliminate the need for additional supervision while focusing on target structures of varying sizes and shapes (Oktay et al., 2018).

While these networks have been created and optimized for medical imaging segmentations, they have found wide-spread application in materials, earth, and life science applications as well. For instance, the skeletal morphology of marine Acantharia has been resolved using synchrotron X-ray nanotomography and U-Net segmentation models (Raja Somu et al., 2023). In geosciences, neural network algorithms have been applied in the tomographic analysis of micro-porous granodiorite rocks (Roslin et al., 2022). Similarly, manufactured aluminum alloys have been visualized using micro-CT, employing U-Net models for porosity and density determinations under tensile stress (Zhang et al., 2023). U-Net models have been employed for studying the morphology of marine bivalve skeletons (Edie et al., 2023). Another study involving deep learning image segmentation revealed nanostructural features in shark vertebral cartilage (Raja Somu et al., 2025). Thus, these models have proven to be viable tools for deep learning micro-CT image segmentations of skeletal systems and biomineralized tissues.

The present study focused on reef-building scleractinian coral species Montastraea cavernosa (M. cavernosa), which has moderate susceptibility to Stony Coral Tissue Loss Disease (SCTLD) (Papke et al., 2023), an aggressive coral disease that emerged in 2014 and has decimated entire coral colonies across Florida and the Caribbean over the last few years (Alvarez-Filip et al., 2022). Another species, Porites astreoides (P. atreoides), which has commonly been used in micro-CT studies, was also included to diversify the datasets used for testing the U-Net models. Herein, we investigated micro-CT reconstructions of M. cavernosa, including healthy and SCTLD-afflicted bulk sections, and P. astreoides to investigate differences in porosity and pore size distributions using deep learning image segmentation algorithms. Micro-CT datasets were acquired in standard absorption mode using a laboratory-source micro-CT scanner. U-Net models were employed and evaluated in a comparative performance study. Upon evaluation, the models were then used to generate labeled image stacks for subsequent skeletal porosity, density, and thickness analyses. The top performing model, Attention U-Net was used to generate labeled image stacks for subsequent porosity analyses. Overall, this work aims to evaluate the performance of deep learning image segmentations compared to traditional manual labeling methods, like histogram-based thresholding, to derive quantitative parameters (e.g., porosity) of healthy and SCTLD-afflicted corals.