Abstract Objectives Usual interstitial pneumonia (UIP) usually indicates that the patient has developed pulmonary fibrosis, suggesting a poor prognosis for the patient. Utilizing machine learning (ML) model to identify UIP based on chest CT scans as early as possible will be beneficial for the diagnosis, treatment and prognosis of the patients, and can also reduce interobserver variability. Methods 723 patients with UIP(n = 260) and non-UIP pattern(n = 463) were included in derivation cohort (n = 346), internal validation cohort (n = 106) and two external validation cohorts (n = 70 and n = 201). Patients of derivation cohort were randomly divided into training set (n = 294) and test set(n = 52). The volume of abnormal CT features was quantified using deep learning software from full CT volumetric images. Then the LASSO regression followed by logistic regression were applied to identify the independent CT features for ML model construction. Finally, five quantitative CT features were calculated valuable to construct ML models and seven mainstream ML models were constructed. Performances of seven models were compared by receiver operating characteristic (ROC) curve, calibration curve, decision curve analysis and confusion matrix. Survival analysis between ML-predicted UIP and human visual UIP were compared by Kaplan-Meier analysis and Cox regression. SHapley Additive exPlanations (SHAP) was used to explain and visualize the clinical application of this model. A flowchart of this study was presented in Figure 1A. Results Five quantitative CT features for ML models include Reticular Volume (right lung), Ground-glass opacities Volume (right lower lung), Honeycomb Volume (right lower lung), Consolidation Volume (left lower lung) and Honeycomb Volume (left lower lung). By multiple comparison of model performances, Logistic Regression (LR) model was identified as the optimum model which showed exceptional discrimination and stability with area under curve (AUC) 0.935 and accuracy 0.904 in testing set, with AUC 0.948 and accuracy 0.84 in internal validation cohort. It also achieved robust discriminatory performance in two external validation cohorts. Survival analysis indicated that ML-predicted UIP has similar survival to those determined by radiologists. SHAP interpretation indicated the importance rank of five CT features and easy clinical applicability for identification of UIP and non-UIP. By SHAP interpretation, two exemplary patients with UIP and non-UIP predicted by LR model were presented in Figure 1(B). Conclusions From this model, by inputting quantitative CT features can easily achieve imaging classification of UIP and non-UIP with barely interobserver bias. This UIP classification model demonstrates excellent discrimination ability and clinical application potential. This abstract is funded by: National Natural Science Foundation of China
Tian et al. (Fri,) studied this question.