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

Abstract

Pulmonary fibrosis is a devastating chronic lung disease characterized by irreparable lung tissue scarring, leading to a progressive loss in lung capacity. A critical component of clinical management is the accurate assessment of lung function decline, often measured via Forced Vital Capacity (FVC). This paper introduces Fibrosis-Net, a deep convolutional neural network (CNN) specifically tailored to predict pulmonary fibrosis progression by fusing chest CT imaging with clinical metadata. Leveraging machine-driven design exploration, the proposed architecture achieves state-of-the-art performance on the OSIC Pulmonary Fibrosis Progression Challenge dataset. Furthermore, explainability-driven validation confirms that the model utilizes clinically relevant visual indicators, such as honeycombing, to inform its predictions.

Introduction

Predicting the progression of pulmonary fibrosis is a significant challenge in respiratory medicine. The disease's trajectory is highly variable, and visual indicators in computed tomography (CT) scans—such as ground-glass opacities or cystic "honeycombing"—can be subtle or mimic other pathologies. Clinicians rely on FVC measurements to track decline, but a reliable automated method to predict future FVC values from initial scans and metadata has remained elusive.

Existing AI approaches often struggle with the high dimensionality of CT data or operate as "black boxes," lacking the transparency required for clinical adoption. Fibrosis-Net addresses these issues through a specialized architecture designed via automated exploration and a focus on clinical interpretability.

Methodology

The development of Fibrosis-Net is centered on three technical pillars: machine-driven design, a tailored architecture for high-dimensional data, and multimodal fusion.

Machine-Driven Design Exploration

Rather than relying on manual architectural tuning, the authors employed a generative synthesis approach. This algorithmic strategy explores a vast space of potential network designs to identify an optimal balance between representational power and computational efficiency. The resulting backbone is optimized specifically for the nuances of CT lung analysis.

The PRPE Architecture

The core of Fibrosis-Net is the Projection-Replication-Projection-Expansion (PRPE) block. This design utilizes lightweight depthwise and pointwise convolutions to process CT stacks efficiently. To further align the model with clinical reality, the researchers implemented a "Lower 55%" heuristic, focusing the network's attention on the lower lung lobes where fibrotic changes are typically most prevalent.

The Fibrosis-Net architecture featuring PRPE blocks and multimodal fusion of CT scans and clinical metadata.

Multimodal Fusion and Explainability

Fibrosis-Net does not rely on imagery alone. It integrates clinical metadata—including age, sex, smoking status, and initial spirometry measurements—late in the architecture to predict the linear rate of FVC change. To ensure the model's decisions are based on pathology rather than image artifacts, the authors utilized GSInquire. This explainability tool audits the decision-making process, verifying that the model focuses on relevant features like honeycombing areas used by radiologists.

Experimental Results

The model was evaluated using the OSIC Pulmonary Fibrosis Progression Challenge dataset, comprising 172 training cases and a dedicated test set. Performance was measured using the Modified Laplace Log Likelihood, a metric that penalizes both prediction error and overconfidence.

Comparison of Fibrosis-Net against the top three winning solutions from the OSIC Pulmonary Fibrosis Progression Challenge.

As shown in the results, Fibrosis-Net achieved a score of -6.8188, significantly outperforming the first-place Kaggle solution (-6.8305). Given the logarithmic nature of the scale, this represents a substantial improvement in predictive reliability and accuracy over previous state-of-the-art models.

Discussion and Conclusion

Fibrosis-Net demonstrates the power of combining automated architectural search with clinical domain knowledge. By focusing on the lower lung regions and incorporating metadata, the model achieves superior predictive performance while maintaining a level of transparency through explainability-driven validation.

Despite these gains, the authors note that Fibrosis-Net is not yet a production-ready clinical tool. The dataset size (approximately 200 patients) remains small for deep learning standards, and further validation in "human-in-the-loop" clinical scenarios is necessary. However, as part of the OpenMedAI initiative, the open-source release of Fibrosis-Net provides a robust foundation for future research into chronic lung disease progression.

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 behaviour 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 Fibrosis-Net exhibits correct decision-making behaviour by leveraging clinically-relevant visual indicators in CT images when making predictions on pulmonary fibrosis progress. While Fibrosis-Net is not yet a production-ready clinical assessment solution, we hope that its release in open source manner will encourage researchers, clinicians, and citizen data scientists alike to leverage and build upon it.
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