Project Summary/Abstract
Progressive fibrosing interstitial lung disease (PF-ILD) is a group of diseases characterized by increasing self-
sustaining fibrosis, progressive worsening of dyspnea, progressive decline in lung function, limited response to
immunomodulatory therapies, and high mortality. Due to the highly variable rates of decline and poor prognosis,
accurate individualized prognostic prediction of patients with PF-ILD is crucial for therapeutic decision making
and management of the patients. However, no formal staging system based on prognosis has been established
for PF-ILD. This is because, despite many attempts, none of the developed existing prognostic biomarkers have
been found to be accurate enough for establishing such a staging system for PF-ILD. A clinically useful staging
system for PF-ILD would enable many important clinical use cases, such as determining the timing and benefits
of the currently available but costly therapies and interventions, identifying patients where treatment can be
safely delayed to avoid potential adverse drug effects and costs, and identifying new therapies in clinical trials.
Quantitative high-resolution computed tomography (HRCT) images have recently emerged as the most
promising approach for providing accurate and reproducible biomarkers in PF-ILD patients, but current HRCT
biomarkers have still yielded only mediocre predictive performances of 64-77% for patients with PF-ILD, as
measured by the concordance index. Thus, there is an unmet clinical need for a prognostic biomarker that would
predict the mortality and disease progression in PF-ILD patients at a high accuracy. Artificial intelligence (AI),
especially deep learning, could be used to realize such a prognostic biomarker. In particular, a conditional
generative adversarial network (cGAN) was recently shown to outperform traditional survival analysis methods
in survival prediction, but there are no such cGAN-based methods to perform prognostic prediction from the
image data of patients. In this project, we propose to develop an unsupervised image-based 3D cGAN model
that would automatically estimate the distribution of the survival time directly from the HRCT images of patients
for prognostic prediction. Our goal is to develop an integrated AI survival prediction model that will combine
existing biomarkers with the image-based 3D cGAN model for performing accurate prognostic prediction in
patients with PF-ILD. We hypothesize that the integrated AI model will yield a high performance (concordance
index of =92%) in predicting the mortality and disease progression in PF-ILD patients. Successful development
of the proposed integrated AI model will significantly improve the accuracy of the current state-of-the-art in the
prognostic prediction of the mortality and disease progression in patients with PF-ILD, thereby ultimately making
it possible to establish a formal staging system for enabling effective management of the patients with PF-ILD.