Segmentation of trabecular bone microdamage in Xray microCT images using a two-step deep learning method.

Area of Science:

• Biomedical Engineering
• Materials Science
• Computational Biology

Background:

• Osteoporosis research benefits from studying bone microdamage under mechanical loading.
• Current methods for bone microdamage evaluation use toxic heavy metal contrast agents and extensive tissue preparation.
• There is a need for advanced, less invasive techniques for microdamage analysis.

Purpose of the Study:

• To evaluate the potential of deep learning models, specifically YOLOv4 and Unet, for detecting and segmenting linear microcracks in trabecular bone.
• To compare the performance of different YOLOv4 model versions for microdamage detection.
• To assess the segmentation accuracy of Unet in microdamage regions.

Main Methods:

• Six bovine bone cylinders were subjected to compression and imaged using microCT.
• Unstained samples were used to train YOLOv4 for microdamage detection and Unet for pixel-level segmentation.
• Performance was evaluated using Intersection over Union (IoU), mean average precision (mAP), and Dice Score.
• Qualitative comparison was made between deep learning and contrast agent approaches.

Main Results:

• The YOLOv4p5 model achieved the best performance in microdamage detection.
• Unet demonstrated effective segmentation, with performance influenced by sample contrast-to-noise ratio.
• Deep learning segmentation focused on crack interiors, while contrast agents highlighted surrounding areas or undamaged regions.

Conclusions:

• The combined YOLOv4 and Unet approach shows promise for microdamage detection and segmentation in trabecular bone.
• The deep learning method is faster than traditional semantic segmentation, despite accumulated errors.
• This technique offers a viable, potentially less toxic alternative for osteoporosis research.