Enhancing trabecular CT scans based on deep learning with multi-strategy fusion.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Radiology

Background:

• Trabecular bone analysis is vital for diagnosing bone diseases like osteoporosis.
• Current 3D CT image restoration methods face significant challenges.
• Limited availability of real human bone Micro-CT datasets hinders research.

Purpose of the Study:

• To develop advanced methods for 3D trabecular computed tomography (CT) image restoration.
• To introduce novel deep learning models for enhanced feature extraction and data fusion.
• To create a new dual-view dataset and validate unsupervised domain adaptation for medical imaging.

Main Methods:

• Development of Cascade-SwinUNETR, a backbone model for single-view 3D CT restoration using deep layer aggregation and Swin-Transformer.
• Introduction of DVSR3D, a dual-view restoration model employing deep feature fusion with attention mechanisms and Autoencoders.
• Curation of a novel dual-view dataset for CT image restoration and implementation of an Unsupervised Domain Adaptation (UDA) method.

Main Results:

• Cascade-SwinUNETR demonstrated strong feature extraction capabilities.
• DVSR3D achieved high performance through effective deep feature fusion.
• The UDA method enabled model adaptation without additional labels, showing potential for clinical applications.
• Validation through downstream medical bone microstructure measurements confirmed the dual-view approach's efficacy.

Conclusions:

• The developed models and dataset significantly advance 3D trabecular CT image restoration.
• Unsupervised Domain Adaptation offers a promising, non-invasive approach for real-world medical applications.
• These contributions enhance trabecular bone analysis, potentially improving clinical outcomes in bone health assessment and diagnosis.