3D tibial HU reconstruction from biplanar X-rays utilizing a hybrid PCA-CNN framework.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• High-resolution Computed Tomography (CT) is essential for bone assessment but limited by high radiation dose, cost, and accessibility.
• Frequent imaging for certain patients exacerbates these limitations.
• Existing methods lack the detail required for precise clinical tasks at reduced doses.

Purpose of the Study:

• To develop and validate a novel hybrid framework for reconstructing 3D tibial CT volumes from biplanar radiographs.
• To assess the feasibility of reducing radiation dose in bone imaging while preserving diagnostic quality.
• To enable detailed internal density distribution analysis from low-dose imaging.

Main Methods:

• A hybrid framework combining statistical intensity modeling (Principal Component Analysis - PCA) with Deep Learning (Convolutional Neural Network - CNN).
• PCA was used to capture intensity variations in a latent space.
• A CNN was trained to regress PCA coefficients directly from biplanar radiographs to reconstruct 3D CT volumes.

Main Results:

• The framework achieved a mean absolute error of 127.17 ± 12.08 HU compared to ground truth CT.
• Structural similarity index was 0.8558 ± 0.0215, and peak signal-to-noise ratio was 21.40 ± 0.78 dB.
• Reconstructed volumes showed interpretable intensity variations, reflecting anatomical differences (e.g., cortical vs. trabecular bone).

Conclusions:

• The proposed hybrid statistical-Deep Learning method can reconstruct 3D bone CT volumes from biplanar radiographs with high fidelity.
• This approach holds potential for substantial radiation dose reduction in bone imaging, aiding frequent imaging scenarios.
• The framework provides a foundation for reduced-dose 3D bone imaging and clinical translation, pending further validation.