Appearance Learning for Image-Based Motion Estimation in Tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Tomographic imaging reconstruction relies on geometric information, vulnerable to patient motion artifacts.
• Patient motion corrupts geometric alignment, leading to inaccurate anatomical structure reconstruction.
• Existing methods struggle to independently address motion-induced geometric deviations.

Purpose of the Study:

• To develop an appearance learning approach for recognizing rigid motion independently of the scanned object.
• To improve the accuracy and robustness of motion artifact correction in tomographic imaging.
• To quantify motion-induced geometric deviations using reprojection error (RPE).

Main Methods:

• A siamese triplet network was trained using a multi-task learning approach.
• The network predicts reprojection error (RPE) for the complete acquisition and along single views.
• Trained on 27 patients with a 21-4-2 split for training, validation, and testing.

Main Results:

• Achieved a residual mean RPE of 0.013mm with an inter-patient standard deviation of 0.022mm, doubling previous accuracy.
• Outperformed two state-of-the-art measures in nine out of twelve experiments in a motion estimation benchmark.
• Demonstrated clinical applicability on a motion-affected clinical dataset.

Conclusions:

• The proposed appearance learning method effectively recognizes and corrects rigid motion artifacts in tomographic imaging.
• This approach offers superior accuracy and robustness compared to existing methods for motion estimation.
• The technique holds significant potential for improving clinical diagnostic image quality in the presence of patient motion.