AI Can Now Make CT Scans Safer and Faster — Without Sacrificing Image Quality
When patients slide into a CT scanner, they absorb more medical radiation than from any other imaging procedure. Roughly 62 percent of all radiation dose that Americans receive from diagnostic imaging comes from CT scans, according to federal data. The risk of harm is small, but it's a number that has made clinicians eager to cut exposure wherever they can — without losing the detail that makes CT so valuable.
The AI Solution
A team at Rensselaer Polytechnic Institute thinks artificial intelligence can do exactly that. In research published this week in Nature Machine Intelligence, engineers led by Ge Wang, a biomedical engineering professor, have developed a deep-learning method that converts low-dose CT images into high-quality scans in a fraction of the time required by current commercial techniques.
The approach relies on what researchers call a modularized neural network — a type of machine learning that progressively strips noise from images while allowing radiologists to steer the process toward the specific diagnosis they're pursuing.
Unlike standard iterative reconstruction, which repeatedly refines images based on predefined rules about physics and image content, this system learns from data and can adapt on the fly.
Testing the Method
To test it, the team gathered low-dose scans from 60 patients — half showing abdominal anatomy, half showing chest — using three different commercial CT platforms already equipped with iterative reconstruction. They ran those same scans through their neural network and then asked three radiologists to score the results. The evaluators rated structural fidelity and noise suppression, the two key markers of whether an image actually tells a doctor what they need to see.
The results were favorable. For abdominal imaging, the AI-generated images scored higher than iterative reconstruction on two of the three scanners and tied with the third. For chest imaging, quality was comparable across all devices. In every comparison, the new method was significantly faster.
The team included collaboration with Mannudeep Kalra, a radiologist at Massachusetts General Hospital and associate professor at Harvard Medical School. Behrouz Shabestari, who directs the National Institute of Biomedical Imaging and Bioengineering's AI program — the body that funded the research — called the work an advance in bringing deep learning to tomographic imaging while improving computational efficiency.
Practical Implications
The implications are practical. Hospitals running older CT hardware could potentially apply this technique without replacing equipment. Wang noted that the method confirms deep learning's viability for producing diagnostic-quality images at lower dosages, sidestepping the time-consuming trade-offs that have historically made dose reduction difficult. The neural network approach avoids the image noise artifacts that iterative methods sometimes introduce.
Based on: Deep Learning–Based Noise Reduction for Low-Dose CT; Ge Wang et al.; Nature Machine Intelligence.