Deep learning methods for automated bone segmentation and biomechanical modelling in zebrafish

Automated segmentation of three-dimensional micro-CT scans is critical for advancing morphometric and biomechanical approaches in musculoskeletal research. In the past decade, the ability to generate and access micro-CT data has increased significantly, producing vast quantities of data that can be prohibitively time-consuming to process manually. Deep learning models generated for specific use cases help to overcome this issue. However, the performance of automated segmentation relative to expert human annotations and their effect on downstream biomechanical analyses remain largely untested. Here, we present the first validated deep learning approach using an attention-augmented 3D U-Net architecture to segment the adult zebrafish (Daino rerio) mandible. The network was trained on a dataset of 47 manually segmented specimens and evaluated using a novel combined Dice-Hausdorff metric that integrates both volumetric and surface accuracy to capture morphology. To assess model performance in practical biomechanical contexts, we compared automated segmentations to those produced by three expert human annotators and constructed finite element (FE) models from each segmentation. Comparisons of Dice-Hausdorff metrics and statistical analyses of FE outputs demonstrates that automated models perform within the range of human segmentations. This framework establishes a generalisable pipeline linking geometric and mechanical validation for biological image analysis.