Framing U-Net via Deep Convolutional Framelets: Application to Sparse-View CT.

Area of Science:

• Medical imaging
• Computational imaging
• Artificial intelligence in healthcare

Background:

• X-ray computed tomography (CT) with sparse projection views reduces radiation dose but causes artifacts with traditional filtered back projection (FBP).
• Deep learning, particularly U-Net architectures, shows promise for sparse-view CT reconstruction but lacks theoretical grounding.
• Artifacts in sparse-view CT limit diagnostic accuracy and necessitate improved reconstruction techniques.

Purpose of the Study:

• To theoretically analyze the limitations of U-Net for sparse-view CT reconstruction.
• To propose novel multi-resolution deep learning schemes inspired by deep convolutional framelet theory.
• To enhance the recovery of high-frequency edges and reduce artifacts in low-dose CT images.

Main Methods:

• Investigated the theoretical limitations of standard U-Net architectures in sparse-view CT.
• Developed and proposed alternative U-Net variants, including dual frame and tight frame U-Nets.
• Validated the proposed methods using extensive experiments on a real patient dataset.

Main Results:

• Demonstrated that standard U-Net architectures have limitations in effectively reconstructing sparse-view CT images.
• Showcased that dual frame and tight frame U-Net variants satisfy the frame condition, enabling better high-frequency edge recovery.
• Achieved superior reconstruction performance compared to existing methods on real patient data.

Conclusions:

• The proposed multi-resolution deep learning schemes, particularly framelet-based U-Net variants, offer significant improvements for sparse-view CT reconstruction.
• These novel architectures provide a theoretically justified approach to overcoming the limitations of traditional methods and standard deep learning models.
• The findings suggest a pathway towards more accurate and reliable low-dose CT imaging for clinical applications.