Deep learning software and revised 2D model to segment bone in micro-CT scans.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Deep learning (DL) struggles with bone segmentation generalization in micro-CT across varied datasets.
• Existing DL models face challenges with data leakage, high memory usage, and limited multi-GPU support.

Purpose of the Study:

• To present BP-2D-03, a revised 2D Bone-Pores segmentation model for improved micro-CT bone analysis.
• To introduce the BONe DL software interface (BONe DLFit, BONe DLPred, BONe IoU) for managing large, diverse datasets.
• To evaluate model robustness, performance across different architectures, and cross-platform reproducibility.

Main Methods:

• Trained BP-2D-03 on a dataset of 20 micro-CT scans from five mammalian species (142,960 image patches).
• Utilized a DL software interface with modules for training, prediction, and evaluation.
• Conducted 5-fold cross-validation, benchmarking experiments on architecture/patch size, and cross-platform consistency tests.

Main Results:

• BP-2D-03 demonstrated stable, high mean Intersection-over-Union (IoU) across seeds, with some variation for atypical scans.
• U-Net and UNet++ architectures with simple convolutional backbones achieved IoU values approaching 0.97.
• Segmentation results were consistent across different hardware, operating systems, and software implementations.

Conclusions:

• The BONe DL software provides a robust baseline for bone segmentation in micro-CT data.
• The developed tools address prior limitations in DL model training and deployment.
• The model and software ensure reproducible and reliable automated bone segmentation across diverse applications.