Abstract Rationale Interstitial lung diseases (ILDs) are a diverse group of lung disorders with variable progression and prognosis. Pulmonary hypertension (PH), a complication of ILD, is associated with increased mortality. With new therapies emerging for ILD-associated PH, PH detection has become increasingly important. Because right heart catheterization, the diagnostic gold standard, is invasive and often unsuitable for severe cases, noninvasive imaging methods are needed. Computed tomography (CT) plays a key role in diagnosis and risk stratification, but conventional visual assessment is limited by subjectivity. Quantitative multicompartment imaging analysis integrating lung parenchyma and pulmonary vasculature on CT may offer a more comprehensive, noninvasive approach to assess disease severity and outcomes. Methods We used CT-derived lung and pulmonary artery radiomic features to improve prognostication and enable noninvasive PH assessment in ILD. Baseline CT scans and clinical data from the Stanford ILD cohort were analyzed for transplant-free survival (n = 1,622) and PH prediction (n = 556 with transthoracic echocardiography). PH was defined as right ventricular systolic pressure greater than 40 mmHg on echocardiography. We used TotalSegmentator V2 for lung and pulmonary artery segmentation. To standardize visualization, lung windowing was applied with a window level (WL) of -600 and window width (WW) of 1,500, while vascular windowing (WL 200, WW 600) was applied for pulmonary artery feature extraction. Contrast phase was labeled using the TotalSegmentator XGBoost classifier, stratifying scans into native, early arterial, late arterial, and portal venous phases. Radiomic features were extracted with Pyradiomics. Nested cross-validation with elastic net penalized Cox models was used to evaluate predictive performance across lung, pulmonary artery, and combined feature sets. Results Lung radiomics models achieved a median concordance index (C-index) of 0.78 for transplant-free survival, outperforming clinical models using age, sex, and forced vital capacity (FVC). Pulmonary artery radiomics alone demonstrated significant predictive value (median C-index 0.71), with further improvement when combined with lung features (median C-index 0.79). For PH prediction, the combined model yielded the highest C-index (up to 0.78), surpassing models based on single compartments (Figure 1). Prognostic features included texture heterogeneity and vascular shape descriptors, suggesting structural remodeling correlates with poor outcomes. Conclusions This study highlights the added value of pulmonary artery radiomics features and supports multicompartment CT analysis as a noninvasive strategy for outcome prediction and PH detection in ILDs, enabling earlier identification of candidates for emerging PH-directed therapies. This abstract is funded by: The Scientific and Technological Research Council of Türkiye
Er et al. (Fri,) studied this question.