Automated Joint Space Detection Improves Bone Segmentation Accuracy.

Area of Science:

• Biomedical Imaging
• Computational Anatomy
• Deep Learning Applications

Background:

• Manual segmentation of complex anatomical structures is labor-intensive and prone to variability.
• Digital image analysis, particularly deep learning (DL), offers automated solutions for landmark detection.
• Existing semi-automated (SA) methods have limitations in accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a supervised deep learning model for automated segmentation of murine paw joints.
• To compare the performance of the DL model against semi-automated methods.
• To assess the model's accuracy in healthy and arthritic mouse models.

Main Methods:

• Supervised analysis of micro-computed tomography (micro-CT) datasets of murine hindpaws and forepaws.
• Developed a 3D U-Net DL model with a ResNet-18 backbone for joint space prediction.
• Incorporated structure enhancement, tensor voting, and output dilation into the segmentation pipeline.
• Utilized wild-type (WT) hindpaw labels as ground truth for training and testing.

Main Results:

• DL model achieved significantly higher segmentation accuracy compared to SA methods in WT mice (e.g., WT male: DL 98.16% vs SA 79.39%).
• Accuracy decreased in tumor necrosis factor transgenic (TNF-Tg) mice with inflammatory-erosive arthritis, correlating with disease severity and age.
• Forepaw segmentation accuracy also declined with increasing arthritis severity in TNF-Tg mice.

Conclusions:

• The supervised DL model substantially outperforms SA approaches for segmenting healthy murine paw joints.
• The model demonstrates potential for investigating complex bone anatomy but shows reduced performance in disease-modified datasets.
• Further development is needed to optimize DL models for complex and pathological anatomical variations.