Data Extrapolation From Learned Prior Images for Truncation Correction in Computed Tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Data truncation in computed tomography (CT) leads to artifacts and missing information, impacting diagnostic accuracy.
• Deep learning shows promise in CT reconstruction but faces challenges with robustness in clinical settings.
• Existing methods for truncation correction often lack the interpretability and reliability needed for widespread clinical adoption.

Purpose of the Study:

• To propose a general plug-and-play (PnP) method for robust CT truncation correction.
• To integrate deep learning-based prior information with conventional image reconstruction algorithms.
• To evaluate the efficacy and robustness of the PnP method using state-of-the-art deep learning models and clinical data.

Main Methods:

• A plug-and-play (PnP) framework was developed, combining data consistency for measured data with learned priors for truncated data.
• Two deep learning models, FBPConvNet and Pix2pixGAN, were integrated into the PnP framework for truncation correction in cone-beam CT.
• The method was tested in noise-free and noisy conditions, with robustness assessed using false lesion cases and quantitative metrics like root-mean-square error (RMSE).

Main Results:

• The PnP method significantly improved robustness and corrected false lesion structures introduced by deep learning models alone.
• In noisy conditions, the PnP approach reduced RMSE inside the field-of-view from 92HU to approximately 30HU for FBPConvNet.
• For Pix2pixGAN, PnP further enhanced image quality, reducing RMSE from 42HU to around 27HU, and demonstrated effectiveness on real clinical head CT data.

Conclusions:

• The proposed plug-and-play (PnP) method offers a robust and interpretable solution for CT truncation correction.
• Integrating deep learning with conventional reconstruction via PnP overcomes limitations of deep learning-only approaches.
• The PnP method shows significant potential for improving clinical CT image quality and diagnostic reliability, as validated on real patient data.