Prognosis prediction for idiopathic pulmonary fibrosis with the 3D generative adversarial network pix2surv3D: improved risk stratification and comparison with a 2D model

Accurate prognosis prediction in idiopathic pulmonary fibrosis (IPF) remains a major clinical challenge due to the disease’s heterogeneous progression and high mortality. In this study, we propose pix2surv3D, a three-dimensional extension of the conditional generative adversarial network (cGAN)-based model pix2surv, designed to predict survival time directly from volumetric chest CT images. Unlike the original 2D model, pix2surv3D employs 3D convolutional layers to capture inter-slice spatial continuity and incorporates regularization techniques including Dropout in the discriminator and label smoothing to improve training stability and generalization. We evaluated the prognostic performance of pix2surv3D using a dataset of 194 IPF patients and compared it with the original pix2surv model. Internal validation was conducted using Efron’s 0.632 bootstrap method with 30 resampling iterations. Results showed that pix2surv3D achieved a comparable Uno’s C-index (0.6824 vs. 0.6947) and a lower Relative Absolute Error (RAE) (0.2538 vs. 0.2675), indicating improved absolute prediction accuracy. In addition, Kaplan–Meier survival curves stratified by the predicted survival times showed more distinct separation between low- and high-risk groups for pix2surv3D, supported by the log-rank test. These findings suggest that the incorporation of 3D spatial features and regularization enhances the model’s risk stratification capability and clinical utility. Our results demonstrate the potential of pix2surv3D as a robust tool for personalized prognosis prediction in interstitial lung disease. Future work includes external validation and expansion to multimodal integration with clinical variables.