Harnessing Radiomics and Explainable AI for the Classification of Usual and Nonspecific Interstitial Pneumonia

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate differentiation between usual interstitial pneumonia (UIP) and nonspecific interstitial pneumonia (NSIP) is critical for effective interstitial lung disease (ILD) management.
• Current diagnostic methods may have limitations in distinguishing these conditions.

Purpose of the Study:

• To evaluate the diagnostic performance of clinical, radiomic, and combined models in classifying UIP versus NSIP using HRCT scans.
• To assess the utility of radiomic features in improving classification accuracy.

Main Methods:

• Retrospective analysis of 105 HRCT scans (60 UIP, 45 NSIP).
• Development of clinical (demographic, pulmonary function) and radiomic models using pyRadiomics.
• Feature selection via recursive feature elimination and model performance evaluation using AUC.
• Interpretation of model predictions using SHapley Additive exPlanations (SHAP).

Main Results:

• The clinical model achieved an AUC of 0.62.
• The radiomic model demonstrated superior performance with an AUC of 0.90 (sensitivity >85%, specificity >85%).
• The combined model yielded an AUC of 0.86.
• SHAP analysis highlighted texture features (GLCM_Idmn, NGTDM_Contrast) as key classifiers.

Conclusions:

• Radiomic features significantly enhance the accuracy of classifying UIP and NSIP compared to clinical data alone.
• Integration of radiomics into clinical workflows holds potential for reducing diagnostic variability and improving ILD diagnosis.
• Further validation with larger, multicenter datasets is recommended.