Deep learning detects sinus treatment changes on CT scans

A radiologist reading a sinus CT must visually estimate how much of each sinus is clouded. Researchers including Stephen M. Humphries at National Jewish Health found that a deep-learning tool detected treatment-related changes more sensitively than that traditional approach in data from two clinical trials.

The tool produces an automated sinus severity score, or SSS. It measures the amount of sinus opacification — the clouding visible on a CT scan — instead of asking clinicians to assign a visual score. The comparison was with the Lund-Mackay system, a widely used method that requires radiologists to judge the degree of opacification.

The target condition, chronic rhinosinusitis with nasal polyps, involves persistent inflammation, swelling and polyps that can block the nasal passages and sinuses. CT imaging is commonly used to assess how extensive the disease is and serves as an important endpoint in clinical trials, but smaller changes can be difficult to capture with a subjective scale.

And concretely? A more sensitive measurement could help researchers identify meaningful responses to treatment and make CT imaging more useful when evaluating new therapies. It could also help clarify how sinus disease changes over time, although the study does not show that the tool is already in routine clinical use.

The limits are clear. The researchers tested the approach on data from two clinical trials involving a single treatment. Additional studies are needed to determine whether it performs as well with other therapies and different mechanisms of action. The findings extend earlier work on automated CT analysis in chronic sinus disease, including research on longitudinal changes in patients with cystic fibrosis.