Predicting pulmonary fibrosis disease progression to enable earlier treatment assessment



DarwinAI leverages Explainable AI (XAI) to quickly develop a high-performing and transparent machine learning model that would rank first in the OSIC Pulmonary Fibrosis Progression Challenge

About The Project

Our team, alongside researchers at the Vision and Image Processing Research Group at The University of Waterloo, and the Waterloo Artificial Intelligence Institute, created Fibrosis-Net: a deep convolutional neural network tailored for the prediction of pulmonary fibrosis progression from chest CT images.

By leveraging explainability in a human-machine collaborative design strategy, the research team achieved state-of-the-art performance for lung decline progression prediction and demonstrated the efficacy of machine-driven design exploration for constructing deep neural network designs tailored for clinical decision support tasks.

weeks to create
when compared to winning solutions in the OSCI Pulmonary Fibrosis Progression Challenge
GB of multimodal data to train Fibrosis-Net model

Difference Makers

  • Machine-driven exploration determined a strong architectural design for CT lung analysis
  • Customized network design optimized for predicting forced vital capacity (FVT) based on clinical inputs
  • Explainability-enabled performance validation to study decision-making behavior
  • Significantly outperformed alternatives (Laplace Log Likelihood score)

Advancing the Fight Against Idiopathic Pulmonary Fibrosis

Pulmonary fibrosis is a devastating chronic lung disease that causes irreparable lung tissue scarring and damage, resulting in progressive loss in lung capacity and has no known cure. Patient outcomes can range from long-term stability to rapid deterioration, but doctors aren’t easily able to tell where an individual may fall on that spectrum. A critical step in the treatment and management of pulmonary fibrosis is the assessment of lung function decline, with computed tomography (CT) imaging being a particularly effective method for determining the extent of lung damage caused by pulmonary fibrosis. In July 2020, the Open Source Imaging Consortium (OSIC) launched an AI code contest to see which hackathon group could best predict a patient’s severity of decline in lung function based on a CT scan of their lungs. We won first prize in the OSIC Pulmonary Fibrosis Progression Challenge and released the model to the public, as part of the OpenMedAI and our Open initiative.

Building a High-Performing, Purpose-Specific Network

Rather than simply treating AI as a tool to be leveraged, our researchers employed a human-machine design strategy. This philosophy reimagines AI as a collaborator that learns from a developer’s needs and that subsequently proposes multiple design approaches with different trade-offs—thereby enabling a rapid and iterative approach to model building.

  1. Data preparation

    To begin the process we prepared a training dataset based upon the CT scans, initial spirometry measurement, and clinical metadata provided by OSIC.

  2. Principled network design prototyping

    We constructed a prototype based on human-driven design principles and best practices. Essentially, the prototype provides the initial scaffolding of the model while leaving final architectural decisions to the machine-driven aspect of the process.

  3. Machine-driven design exploration

    Given data and human-defined requirements, our platform guides a design exploration that learns and identifies the optimal macro-and micro-architectures for predicting forced vital capacity based on a CT scan.

  4. Machine-guided design audit

    We leverage explainability to audit the Fibrosis-Net model—validating that Fibrosis-Net is driven by correct, clinically relevant decision-making behaviour when making predictions of pulmonary fibrosis progression, similar to those leveraged by clinicians.

Beyond Fibrosis-Net

Based on both quantitative and qualitative results, it was demonstrated that Fibrosis-Net can not only make FVC predictions at a higher level of accuracy than state-of-the-art methods, but also do it in a more trustworthy, validated manner that leverages clinically relevant visual indicators within the CT images of a pulmonary fibrosis patient.

Since releasing the model to the public, as part of the OpenMedAI and DarwinAI Open initiative, our team has used the Fibrosis-Net model within the pharmaceutical industry to help discover patients for Pulmonary Fibrosis treatments.

These activities were aided by Amazon Web Services (AWS), who facilitated an introduction to an AWS enterprise customer who could partner with us on a pilot.