A residual dense network assisted sparse view reconstruction for breast computed tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Dedicated breast computed tomography (CT) offers improved imaging but involves radiation exposure.
• Reducing radiation dose in breast CT is crucial for patient safety and widespread adoption.
• Current reconstruction methods for sparse-view data often compromise image quality.

Purpose of the Study:

• To develop and evaluate a deep learning framework for radiation dose reduction in breast CT using sparse-view acquisition.
• To investigate the efficacy of a multi-slice residual dense network (MS-RDN) for reconstructing high-quality breast CT images from limited projection data.

Main Methods:

• A framework combining 3D sparse-view cone-beam acquisition with a multi-slice residual dense network (MS-RDN) was proposed.
• Full-scan (300 views) breast CT datasets from 34 women were used as reference, reconstructed with the FDK algorithm.
• Sparse-view (100 views) data were reconstructed using FDK and served as input for the MS-RDN, trained with reference reconstructions as labels.

Main Results:

• The proposed MS-RDN demonstrated superior quantitative and visual performance compared to conventional compressed sensing and other deep learning methods.
• The deep learning approach successfully reconstructed high-quality breast CT images from significantly reduced projection views.
• The MS-RDN achieved image quality comparable to or better than FDK reconstructions from full-scan data.

Conclusions:

• The developed deep learning framework effectively enables low-dose breast CT imaging.
• This approach holds significant potential for reducing radiation exposure in mammography and breast cancer screening.
• The MS-RDN reconstruction method offers a promising solution for improving the diagnostic accuracy of sparse-view breast CT.