B23-12 A Machine Learning-based Radiomics Model for Differentiating Idiopathic Pulmonary Fibrosis and Fibrotic Hypersensitivity Pneumonitis Thorough Pyradiomics

Abstract Rationale Idiopathic pulmonary fibrosis (IPF) and hypersensitivity pneumonitis (HP) are fibrosing interstitial lung diseases that often present with similar clinical and radiologic features, making accurate differentiation challenging. No definitive noninvasive biomarkers exist for distinguishing these conditions, yet accurate distinction is crucial as IPF and fibrotic HP have different management strategies. Recently the radiomics features of the chest field were extracted using Pyradiomics which is a comprehensive open platform to enable processing and extraction of radiomics features from whole chest imaging data using engineered algorithms.We aimed to evaluate whether computed tomography (CT) radiomic features extracted from segmented lung CT images using the PyRadiomics platform, combined with machine learning (Lasso regression), could improve diagnostic discrimination between IPF and fibrotic HP. Methods We retrospectively analyzed 180 biopsy-confirmed cases (IPF, n = 87; CHP, n = 93) diagnosed with multidisciplinary discussion. Clinical variables were collected, including environmental exposure history, bronchoalveolar lavage (BAL) lymphocyte percentage, serologic markers (e.g., serum KL-6), and pulmonary function tests. Radiomic features (shape, intensity histogram, and texture metrics such as gray-level co-occurrence matrix GLCM features) were extracted from segmented high-resolution lung CT scans. Three Lasso logistic regression models were constructed: Model 1 using clinical data alone, Model 2 using radiomic features alone, and Model 3 using combined clinical plus radiomic inputs. Model performance was assessed by area under the ROC curve (AUC), accuracy, sensitivity, and specificity using an internal train/test split. Results IPF group had more male patients than fibrotic HP group (p = 0.003), and the fibrotic HP group had a significantly higher percentage of down quilt use (p = 0.006) and lymphocyte ratio in bronchoalveolar lavage (BAL)than the IPF group (p 0.001). Fibrotic HP cases also more often showed mid-upper lung fibrosis with peribronchovascular distribution on CT, whereas IPF cases showed basal-predominant honeycombing. Lasso regression identified key discriminatory features, including radiomic features like GLCM Difference Variance and Sphericity, and clinical variables like BAL lymphocyte percentage and serum KL-6. The radiomics-only model achieved a test AUC of ∼0.73 and the clinical-only model AUC ∼0.76. The combined model (Model 3) performed best on the test data, with AUC 0.85, accuracy 78%, sensitivity 89%, and specificity 69%, outperforming Models 1 and 2. Conclusion CT radiomic analysis with machine learning improved differentiation between IPF and fibrotic HP. Combining quantitative imaging biomarkers with clinical data enhanced diagnostic accuracy compared to either modality alone. The results suggest that the Pyradiomics has the potential diagnostic method to distinguish between IPF and fibrotic HP without conducting invasive methods. This abstract is funded by: the Ministry of Health, Labour and Welfare