Developing prognostic deep learning models in IPF using OSIC dataset.

<bold>Motivation:</bold> Prognostic models that accurately predict idiopathic pulmonary fibrosis (IPF) progression are needed to improve clinical trial design. Deep learning (DL) models that incorporate high-resolution computed tomography (HRCT) to predict forced vital capacity (FVC) decline may be promising. <bold>Objectives:</bold> To develop and compare 2.5D and 3D HRCT-based DL models against a benchmark model of Gender-Age-Physiology (GAP) features. <bold>Methods:</bold> Retrospective analysis was performed on 459 IPF subjects with baseline HRCT from the Open Source Image Consortium (OSIC). IPF progression was defined as annualized FVC decline (L/yr) and computed as the slope of a linear fit of available temporal FVC measurements. 2.5D (10 randomly selected axial slices) and 3D (volumetric lung scans) DL models using ResNet and DenseNet backbones were trained to predict FVC decline and compared to a benchmark linear model using baseline GAP features. 2.5D image-only and multi-modal (image + GAP features) models were developed. The dataset was split into training (n = 416) and holdout (n = 43). The performance was evaluated on the holdout set using the squared Pearson correlation coefficient (r²). <bold>Results:</bold> The 2.5D multi-modal model had the best performance on the holdout set (r² = 0.20). In comparison, the 2.5D and 3D models using image-only showed an r² of 0.15 and 0.10, respectively on the holdout set. All DL models outperformed the benchmark model (r² = 0.01). <bold>Conclusion:</bold> Incorporating HRCT into DL models improves the prognostication of IPF progression over conventional baseline GAP features supporting potential use in future clinical trials. Further optimization of models include: exploring newer DL architectures and increasing HRCT data.