AI-powered ILD quantification

In a collaboration between the Department of Imaging and Pathology at the KU Leuven and, Radiomics Bio, Flore Belmans and colleagues developed and validated an AI-powered model for CT-based quantification of interstitial lung disease (ILD) patterns in scleroderma (SSc)–associated ILD patients. The objective of the study, presented as a late breaking abstract at ERS 2025, was twofold: (1) build a robust segmentation model for ILD patterns and (2) assess the clinical relevance of these imaging biomarkers in disease progression.

In the first stage, over 8,000 axial CT slices were manually annotated by a team of expert radiologists, covering four key ILD patterns: reticulation, ground-glass opacities, consolidation, and honeycombing. The AI model trained on these annotations was then validated on an external dataset, demonstrating accurate detection of ILD and a strong concordance between automatically quantified volumes and the manually derived ‘ground truth’. This confirms that the model is able to reliability extract quantitative features from CT scans.

In the second stage, the model was applied to a cohort of 82 SSc-ILD patients, investigating whether automatically quantified ILD volumes were associated with pulmonary function tests (PFTs) and with clinical diagnosis of progression. The model revealed a statistically significant difference in ILD volume between progressive and non-progressive patients. Furthermore, multivariable logistic regression showed that reticulation and honeycombing volumes alone yielded an AUC of 0.8 for identifying progression, underscoring their potential as predictive imaging biomarkers.

In conclusion, the presented AI-powered model accurately quantifies ILD patterns and provides clinically relevant biomarkers for SSc-ILD progression. Future directions include extending this approach to other ILD subtypes and exploring its application in oncology, particularly for the quantification of immunotherapy- or radiotherapy-related ILD.