Abstract
Deep learning (DL) enables automated bone segmentation in micro-CT datasets but can struggle to generalize across developmental stages, anatomical regions, and imaging conditions. We present BP-2D-03, which is a revised 2D Bone-Pores segmentation model. It was fitted to a dataset comprising 20 micro-CT scans spanning five mammalian species and 142,960 image patches. To manage the substantially larger and more varied dataset, we developed a DL software interface with modules for training ("BONe DLFit"), prediction ("BONe DLPred"), and evaluation ("BONe IoU"). These tools resolve prior issues such as slice-level data leakage, high memory usage, and limited multi-GPU support. Model performance was evaluated through three analyses. First, 5-fold cross-validation with three seeds per fold evaluated baseline robustness and stability. The model showed generally high mean Intersection-over-Union (IoU) with minimal variation across seeds, but performance varied more across folds related to differences in scan composition. These findings show that the baseline model is stable overall but that predictivity can decline for atypical scans. Second, 30 benchmarking experiments tested how model architecture, encoder backbone, and patch size influence segmentation IoU and computational efficiency. U-Net and UNet++ architectures with simple convolutional backbones (e.g., ResNet-18) achieved the highest IoU values, approaching 0.97. Third, cross-platform experiments confirmed that results are consistent across hardware configurations, operating systems, and implementations (Avizo 3D and standalone). Together, these analyses demonstrate that the BONe DL software delivers robust baseline performance and reproducible results across platforms.