CT feature combinations improve identification of interstitial lung disease patterns

Specific combinations of CT imaging features rather than individual findings alone may improve the accuracy of identifying radiologic patterns in interstitial lung disease (ILD), according to a new study by researchers in the department of medicine at the University of British Columbia in Vancouver, Canada.

The study, published in Radiology last month, highlights how radiologists integrate multiple imaging characteristics when classifying ILD and suggests opportunities to refine diagnostic approaches using more structured models.

Lead author Daniel-Costin Marinescu, M.D., M.H.Sc., a respirologist at Vancouver General Hospital and Centre for Lung Health, in Vancouver, along with Cameron J. Hague, M.D., and colleagues from a number of institutions, including Providence Health Care and the University of British Columbia, used data from the Canadian Registry for Pulmonary Fibrosis (CARE-PF) to examine how individual CT features contribute to overall pattern recognition in ILD.

The researchers noted high-resolution CT plays a central role in diagnosing and classifying ILD, guiding treatment decisions and determining whether invasive procedures such as lung biopsy are necessary. However, interpretation can be complex, as many imaging features overlap across disease subtypes, including usual interstitial pneumonia (UIP), nonspecific interstitial pneumonia (NSIP) and fibrotic hypersensitivity pneumonitis.

To better understand how radiologists synthesize imaging findings, the investigators applied a multinomial statistical model to evaluate associations between individual CT features and final radiologic diagnoses. Unlike traditional approaches that examine features in isolation, this model allowed the researchers to assess how multiple findings interact simultaneously, more closely reflecting real-world clinical reasoning.

The research revealed that radiologists rely heavily on both the distribution of disease and the presence of distinctive imaging features when determining ILD patterns. Certain findings, such as honeycombing, traction bronchiectasis and reticulation, were strongly associated with specific diagnostic categories, but their significance depended on how they were combined with other features and where they were located within the lung.

For example, the presence of subpleural and basal-predominant abnormalities was an important factor in identifying UIP, while alternative distributions and additional features could shift interpretation toward other ILD subtypes. The authors noted this reinforces the concept that pattern recognition in ILD is multidimensional and cannot be reduced to single imaging markers.

The study also highlighted the limitations of current guideline-based approaches, which often categorize imaging findings into discrete patterns without fully accounting for overlapping or mixed features. By contrast, the multinomial model captured the probabilistic nature of ILD diagnosis, where multiple features contribute varying degrees of influence.

“Our findings suggest that radiologists emphasize disease distribution and distinctive features when identifying patterns,” the authors wrote, noting that integrating these elements into a unified model more accurately reflects clinical decision-making.

These results have potential implications for both clinical practice and research. More structured approaches to CT interpretation could improve diagnostic consistency among radiologists and support multidisciplinary discussions, which are critical for accurate ILD diagnosis.

Additionally, quantitative or model-based tools could help standardize reporting and reduce variability in interpretation across centers.

The study’s findings could also inform the development of artificial intelligence and machine learning tools designed to assist in ILD classification. By incorporating combinations of imaging features and their relative importance, such systems could more closely replicate expert-level reasoning and improve diagnostic performance.

The authors acknowledged limitations, including the observational design and reliance on registry data, which may introduce variability in imaging acquisition and interpretation. Additionally, while the model provides insight into associations between features and diagnoses, it does not replace the need for clinical correlation or multidisciplinary evaluation.

Still, the study adds to a growing body of evidence emphasizing the importance of comprehensive imaging assessment in ILD. As diagnostic frameworks evolve, integrating multiple imaging features into cohesive models may help clinicians better characterize disease patterns and guide management decisions.