A study on early diagnosis for fracture non-union prediction using deep learning and bone morphometric parameters

Area of Science:

• Biomedical Engineering
• Radiology
• Orthopedics

Background:

• Early diagnosis of fracture non-union is critical for effective treatment planning.
• Current diagnostic methods lack efficiency, and studies on bone morphometric parameters are limited.
• This research addresses the need for advanced diagnostic tools in fracture healing assessment.

Purpose of the Study:

• To create a high-resolution micro-CT image dataset of rat fracture healing.
• To develop a deep learning algorithm (VM-TE-UNet) for accurate fracture segmentation.
• To establish an early diagnostic model for fracture non-union using bone morphometric parameters.

Main Methods:

• Micro-CT imaging of rat fracture models at multiple healing stages (days 1-35).
• Annotation of fracture lesions to build a comprehensive image dataset.
• Implementation and validation of the Vision Mamba Triplet Attention and Edge Feature Decoupling Module UNet (VM-TE-UNet) for segmentation.
• Extraction of bone morphometric parameters for diagnostic model development.

Main Results:

• A dataset of 2,448 micro-CT images was established.
• The VM-TE-UNet achieved a Dice Similarity Coefficient of 0.809 for fracture segmentation, outperforming baseline models.
• The early diagnosis model demonstrated high accuracy (AUC of 0.995) by day 14, identifying non-union during the soft-callus phase.
• Significant differences in bone morphometric parameters were found between healing and non-union groups (p < 0.05).

Conclusions:

• The VM-TE-UNet effectively segments fracture areas, facilitating the extraction of diagnostic parameters.
• The developed model enables early prediction of fracture non-union (as early as 14 days), aiding in timely clinical intervention.
• This approach offers a valuable reference for clinical non-union prediction and early treatment strategies, particularly in cases of compromised blood supply.