CO:05:3 | Radiomics and machine learning for differentiating rheumatoid arthritis-associated interstitial lung disease and idiopathic pulmonary fibrosis with usual interstitial pneumonia pattern

Vincenzo Venerito¹, Chiara Nani², Andreina Manfredi³, Marco Fornaro¹, Giuseppe Lopalco¹, Florenzo Iannone¹, Cecilia Burattini², Marco Sebastiani⁴. | ¹Rheumatology Unit, University of Bari; ²Respiratory Unit, Guglielmo da Saliceto Civil Hospital, Piacenza; ³Rheumatology Unit, Azienda Unità Sanitaria Locale-IRCCS di Reggio Emilia; ⁴Rheumatology Unit, AUSL Piacenza, University of Parma, Italy

Background. While Idiopathic Pulmonary Fibrosis (IPF) and Rheumatoid Arthritis-associated Interstitial Lung Disease (RA-ILD) can manifest identical Usual Interstitial Pneumonia (UIP) patterns on HRCT, their underlying histopathology reveals fundamentally different disease processes—IPF driven by aberrant fibroblast proliferation and excessive collagen deposition versus RA-ILD's autoimmune-mediated inflammation. This pathological divergence translates to dramatically different prognoses and treatment responses, yet current imaging interpretation cannot reliably capture these microscopic differences. Radiomics offers a revolutionary approach: extracting hundreds of quantitative features that may encode histopathological signatures within standard HRCT images. Combined with machine learning, this technology might unveil tissue-level distinctions invisible to visual analysis, potentially transforming a diagnostic challenge into a data-driven classification. Our aim was to develop a radiomic-based framework for distinguishing IPF from RA-ILD in UIP-pattern patients.

Methods. We conducted a retrospective analysis of patients with confirmed UIP pattern diagnosed in tertiary centres. Whole-lung volumes were semi-automatically segmented using 3DSlicer (v5.4.0, https://tinyurl.com/mr339kmm). 120 radiomic features were extracted via PyRadiomics (v3.0.1). Features underwent z-score standardization and multicollinearity assessment Two distinct XGBoost models were developed to address different clinical scenarios: Standard Model: Leveraged all radiomic features to maximize discriminative performance. - Extent-Aware Model: Excluded volume-related features and incorporated mesh volume as a covariate to control for disease severity, ensuring classification based on intrinsic tissue characteristics rather than disease burden. Model performance was evaluated using 5-fold cross-validation with area under the receiver operating characteristic curve (AUC) as the primary metric.

Results. The study included 72 patients: 42 with IPF (male 73.8%, median age 78, IQR 8) and 30 with RA-ILD (female, 63.3%, median age 72. IQR 12.5). Standard Model achieved near-perfect discrimination (AUC 0.99±0.01), driven by first-order intensity features: Image-original_Mean (overall lung density); Image-original_Maximum/Minimum (attenuation extremes); Extent-Aware Model maintained robust performance (AUC 0.77±0.09) using texture-based features: original_glcm_Id (texture homogeneity); original_firstorder_10Percentile (lower intensity distribution); original_glszm_ZonePercentage (homogeneous zone proportion) IPF median survival 5.5 months versus 29.5 months for RA-ILD.

Conclusions. Radiomics with machine learning demonstrates exceptional potential for differentiating IPF from RA-ILD in UIP-pattern patients. The near-perfect accuracy suggests quantitative biomarkers capture disease-specific signatures beyond visual assessment. Crucially, the extent-aware model's success indicates fundamental tissue architecture differences between conditions, independent of disease severity. Implementing radiomic analysis could enable earlier diagnosis and timely disease-specific therapy.