A truth-based primal-dual learning approach to reconstruct CT images utilizing the virtual imaging trial platform

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Computed tomography (CT) image reconstruction is an ill-posed inverse problem requiring regularization for accuracy.
• Data-driven methods for CT reconstruction need large datasets with ground truth, which are often difficult to obtain.
• Virtual imaging trials offer a solution for generating high-fidelity simulated projection data and ground truth.

Purpose of the Study:

• To develop a novel strategy for CT image reconstruction using a virtual imaging trial (VIT) platform.
• To leverage a learned primal-dual deep neural network (LPD-DNN) for accurate image reconstruction.
• To demonstrate the effectiveness of VIT-generated ground truth for training deep learning models.

Main Methods:

• A learned primal-dual deep neural network (LPD-DNN) was employed, incorporating the forward model and its adjoint.
• A virtual imaging trial (VIT) platform, utilizing XCAT computational models and the DukeSim simulator, generated noise-free projection data and ground truth.
• Noisy sinogram data and corresponding linear attenuation coefficients were used for training the LPD-DNN.

Main Results:

• The LPD-DNN achieved a 12% normalized root mean square error (NRMSE) compared to ground truth.
• Reconstructed images showed a peak signal-to-noise ratio (PSNR) of 32 dB and a structural similarity index (SSIM) of 96%.
• These results significantly outperformed standard filtered-back projection (FBP) reconstruction.

Conclusions:

• The proposed VIT-based LPD-DNN strategy enables accurate CT image reconstruction.
• This approach overcomes the limitations of traditional methods by providing accessible ground truth data.
• The method demonstrates superior performance in image quality and accuracy compared to FBP reconstruction.