Fibrosis-Net: A Tailored Deep Convolutional Neural Network Design for Prediction of Pulmonary Fibrosis Progression From Chest CT Images

Alexander Wong, Jack Lu, Adam Dorfman, Paul McInnis, Mahmoud Famouri, Daniel Manary, James Ren Hou Lee, Michael Lynch

TL;DR

This paper introduces Fibrosis-Net, a novel deep convolutional neural network designed to predict the progression of pulmonary fibrosis using chest CT scans. By employing a machine-driven design exploration and an explainability-driven validation strategy, Fibrosis-Net achieves state-of-the-art performance and demonstrates clinically relevant decision-making, offering a promising tool for computer-aided clinical decision support.

What Problem Does This Paper Address?

Pulmonary fibrosis is a severe, chronic lung disease characterized by irreversible scarring and progressive loss of lung function, with no known cure. Accurately assessing and predicting the rate of this decline is crucial for effective patient management, which often involves oxygen therapy, pulmonary rehabilitation, or pharmacological interventions. While CT imaging is a powerful tool for visualizing lung damage, predicting the progression of fibrosis remains a significant challenge due to the variability in disease presentation and the limitations of traditional spirometry tests in capturing underlying disease mechanisms. Existing methods struggle to accurately forecast future lung function decline, highlighting the need for more sophisticated predictive models.

Methods

The core of this work is Fibrosis-Net, a deep convolutional neural network (CNN) architecture. Its design was guided by a machine-driven design exploration strategy, which iteratively optimized the network architecture based on performance metrics and constraints. This process led to a tailored design that leverages depthwise and pointwise convolutions, along with projection-replication-projection-expansion (PRPE) blocks, to balance representational capacity with computational efficiency.

A key aspect of their approach is the visual-clinical feature fusion, where learned visual features from CT scans are combined with clinical metadata (age, sex, smoking status, spirometry measurements) at a later stage in the network. This allows for a more informed prediction of future Forced Vital Capacity (FVC).

To ensure the model's predictions are trustworthy and clinically relevant, the authors employed an explainability-driven performance validation strategy using GSInquire. This method helps to identify the specific visual features in CT images that Fibrosis-Net uses to make its predictions, verifying that the model is not relying on spurious correlations or artifacts.

Datasets

The study utilized a patient cohort from the OSIC Pulmonary Fibrosis Progression Challenge (OSIC, 2020). This dataset includes chest CT scans, FVC measurements over time, and associated clinical metadata. The authors used 172 patient cases for training and approximately 28 for testing. This cohort is noted as the largest publicly available dataset for this specific challenge.

Key Insights and Ideas

  • Machine-Driven Design for Specialization: The paper demonstrates the power of machine-driven design exploration in creating highly specialized CNN architectures that outperform general-purpose models for specific clinical tasks like pulmonary fibrosis progression prediction.
  • Bridging Visual and Clinical Data: Effectively fusing CT scan features with clinical metadata is crucial for accurate FVC prediction, suggesting that a holistic view of patient data is more powerful than relying on imaging alone.
  • Explainability as a Trust Builder: Using explainability methods like GSInquire is not just for debugging but is essential for building trust and facilitating the adoption of AI in clinical settings, by ensuring predictions are based on clinically meaningful indicators.
  • Open-Source for Acceleration: The release of Fibrosis-Net as an open-source tool is a significant contribution, aiming to accelerate research and development in AI for medical applications.

Key Results

  • State-of-the-Art Performance: Fibrosis-Net achieved a significantly higher modified Laplace Log Likelihood score than the winning solutions of the OSIC Pulmonary Fibrosis Progression Challenge, indicating superior performance in predicting lung function decline. Specifically, it outperformed the Kaggle 1st place winner by 0.0117 on this logarithmic scale.
  • Computational Efficiency: The backbone architecture of Fibrosis-Net is over 22x less complex (fewer parameters) and over 10x more computationally efficient (faster inference speed) compared to the Kaggle 1st place winning solution, making it highly suitable for resource-constrained clinical environments.
  • Clinically Relevant Explanations: Visualizations from GSInquire confirmed that Fibrosis-Net leverages clinically relevant indicators, such as the presence and extent of honeycombing, to make its FVC predictions, mirroring how clinicians would interpret such scans.

Limitations and Future Work

The authors acknowledge that Fibrosis-Net is not yet a production-ready clinical assessment solution. They emphasize that its predictions should be used to aid clinicians, not replace their judgment, advocating for a "human-in-the-loop" approach.

Future work includes:

  • Exploring the Fibrosis-Net strategy for other pulmonary conditions like chronic obstructive pulmonary disease (COPD) and pulmonary hypertension.
  • Conducting deeper exploration and analysis using explainability methods on a larger corpus of patient cases to gain further clinical and model insights.

Summary updated

Abstract

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. 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. Motivated by this, we introduce Fibrosis-Net, a deep convolutional neural network design tailored for the prediction of pulmonary fibrosis progression from chest CT images. More specifically, machine-driven design exploration was leveraged to determine a strong architectural design for CT lung analysis, upon which we build a customized network design tailored for predicting forced vital capacity (FVC) based on a patient’s CT scan, initial spirometry measurement, and clinical metadata. Finally, we leverage an explainability-driven performance validation strategy to study the decision-making behavior of Fibrosis-Net as to verify that predictions are based on relevant visual indicators in CT images. Experiments using a patient cohort from the OSIC Pulmonary Fibrosis Progression Challenge showed that the proposed Fibrosis-Net is able to achieve a significantly higher modified Laplace Log Likelihood score than the winning solutions on the challenge. Furthermore, explainability-driven performance validation demonstrated that the proposed Fib...
BibTeX
@article{wong2021fibrosis,
  author = {Alexander Wong and Jack Lu and Adam Dorfman and Paul McInnis and Mahmoud Famouri and Daniel Manary and James Ren Hou Lee and Michael Lynch},
  title = {{Fibrosis-Net: A Tailored Deep Convolutional Neural Network Design for Prediction of Pulmonary Fibrosis Progression From Chest CT Images}},
  journal = {Frontiers Artif. Intell.},
  year = {2021},
  doi = {10.3389/frai.2021.764047},
}
Figure