Deep learning CT model superior to state-of-the-art methods

More than 80 million CT scans were performed in 2015 in the U.S. While low-dose CT (LDCT) produces more noise than its traditional counterpart, which can limit radiation exposure, the modality does sacrifice diagnostic ability.

A University of Saskatchewan team has created a deep learning technique that demonstrated enhanced de-noising capabilities in LDCT imaging, resulting in little resolution loss and better performance, according to a study published in the Journal of Digital Imaging.

The artificial intelligence (AI) algorithm known as the sharpness-aware generative adversarial network (SAGAN) was applied to both simulated CT image data and real data images.

In the simulated group, 239 normal-dose CT images were downloaded from the National Cancer Imaging Archive (NCIA). To test the algorithm in real data sets, 850 CT scans of a deceased animal model were acquired, with 708 of those used for training the algorithm and 142 for testing.

The SAGAN method recovered the low-contrast zoom regions much sharper than other methods tested in the study. It also recovered small details more effectively, producing clearer images compared to traditional methods tested in the research.

“The proposed SAGAN achieves improved performance in the quantitative assessment and the visual results are more appealing than the tested competitors,” wrote corresponding author Xin Yi, with the University of Saskatchewan’s College of Medicine in Canada, and colleagues.

Authors acknowledge all deep-learning methods, including the ones tested in this study, need to be trained against a specific dosage level. While the team did train their method on clinical patient data, it was mostly geared to analyze the quality of de-noised images. The performance of these algorithms still needs to be tested in clinical practice.