Iterative Image Reconstruction with Under-Sampled Data Assisted by a Neural Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Iterative algorithms are standard for under-sampled image reconstruction, minimizing objective functions with data fidelity and Bayesian terms.
• Total Variation (TV) norm is a common Bayesian term used in image reconstruction.

Purpose of the Study:

• To introduce a novel Bayesian term for image reconstruction.
• To utilize a neural network classifier to identify and penalize artifacts in reconstructed images.

Main Methods:

• A neural network classifier was trained on patient images reconstructed from complete and incomplete datasets.
• The neural network generates a "CNN score" representing artifact severity.
• This CNN score is incorporated as an additional Bayesian term in iterative reconstruction algorithms.

Main Results:

• Patient studies demonstrated a strong correlation between the CNN score and the severity of artifacts.
• The CNN score effectively quantifies artifacts caused by incomplete data.

Conclusions:

• Neural networks can extract image features indicative of incomplete measurements and quantify them as a CNN score.
• Minimizing the CNN score in iterative reconstruction algorithms can suppress artifacts in the final image.