Cycloidal CT with CNN-based sinogram completion and in-scan generation of training data.

Area of Science:

• Medical Imaging
• Computational Imaging
• Machine Learning

Background:

• X-ray computed tomography (CT) resolution is limited by source and detector characteristics.
• Mask-based CT overcomes resolution limits but requires extensive data acquisition.
• Cycloidal CT offers faster scans and lower doses but produces incomplete data.

Purpose of the Study:

• To reconstruct high-quality images from incomplete cycloidal CT data.
• To introduce a novel machine learning-based training approach for data acquisition.
• To demonstrate the effectiveness of cycloidal CT combined with machine learning for high-resolution imaging.

Main Methods:

• Application of a Mixed Scale Dense (MS-D) convolutional neural network (CNN) for sinogram restoration.
• Development of a new training strategy using data acquired during each scan.
• Validation with simulated and real-world datasets.

Main Results:

• Successful reconstruction of high-quality images from incomplete cycloidal CT data.
• Demonstration of accurate high-resolution imaging using the MS-D CNN.
• Validation of the novel training approach, reducing reliance on pre-existing datasets.

Conclusions:

• Cycloidal CT combined with MS-D CNN-based machine learning enables accurate, high-resolution imaging.
• The proposed training method simplifies data acquisition for cycloidal CT.
• This approach significantly reduces radiation dose and scan time in medical imaging.