Accurate and robust sparse-view angle CT image reconstruction using deep learning and prior image constrained...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Sparse-view CT reconstruction is challenging, especially for dynamic acquisitions.
• Current deep learning methods struggle with patient-specific accuracy and cohort generalizability.

Purpose of the Study:

• To address limitations in deep learning-based sparse-view CT reconstruction.
• To improve both individual patient accuracy and generalizability of reconstruction methods.

Main Methods:

• Developed a hybrid approach combining deep learning with prior image constrained compressed sensing (DL-PICCS).
• Initial reconstruction via filtered backprojection (FBP), followed by deep learning for artifact reduction.
• Deep learning output served as the prior image for PICCS, with optional secondary network for noise reduction.

Main Results:

• DL-PICCS significantly improved quantitative reconstruction accuracy compared to existing deep learning and CS methods.
• Effectively eliminated false positives (lesion-like structures) and false negatives (missing anatomy).
• Enhanced generalizability of deep learning schemes by relaxing working conditions.

Conclusions:

• DL-PICCS offers a promising solution for personalized CT image reconstruction.
• Achieves improved accuracy and enhanced generalizability in sparse-view CT.