CT Image Segmentation for Inflamed and Fibrotic Lungs Using a Multi-Resolution Convolutional Neural Network

Sarah E. Gerard, Jacob Herrmann, Yi Xin, Kevin T. Martin, Emanuele Rezoagli, Davide Ippolito, Giacomo Bellani, Maurizio Cereda, Junfeng Guo, Eric A. Hoffman, David W. Kaczka, Joseph M. ReinhardtarXiv 2020 · arXiv

Introduction

Accurate lung segmentation—the precise identification of lung tissue boundaries—is a fundamental prerequisite for quantitative analysis of chest CT scans. However, standard segmentation algorithms frequently fail when lungs are diseased. In conditions such as COVID-19 or Acute Respiratory Distress Syndrome (ARDS), lungs fill with fluid or develop thick fibrous tissue, resulting in consolidation and ground-glass opacities. To traditional computer vision algorithms, these dense "white-out" regions appear more similar to the chest wall or heart than to healthy, air-filled lung tissue. Furthermore, training robust deep learning models typically requires massive amounts of expert-labeled data for every specific disease, a resource that is often unavailable during the onset of a new pandemic.

Methods: Polymorphic Training

To address these challenges, researchers developed a polymorphic training approach. This strategy leverages "partial" or "hierarchical" labels from disparate datasets to teach a single neural network multiple tasks simultaneously. The training utilized two primary data sources:

  1. Human COPD Data: Provides high-specificity labels (distinguishing left vs. right lung) but lacks high-density "injured" phenotypes.
  2. Animal ARDS Data: Provides low-specificity labels (general "lung" vs. "background") but contains the critical high-density abnormalities missing from human datasets.
The polymorphic training architecture uses an auxiliary layer to accommodate datasets with different label specificities.

The core architecture, a multi-resolution CNN (LungNet/Seg3DNet), incorporates an auxiliary layer during training. This layer sums the predicted left and right lung channels into a single "total lung" channel. This mechanism allows the network to be supervised by both specific human labels and nonspecific animal labels within the same training loop, effectively bridging the gap between species and label types.

Key Insights: Phenotype over Species

A central insight of this work is that a network does not need to see "COVID-19" images to segment them accurately; it simply needs to encounter the phenotype of consolidation. By using animal models of lung injury, the researchers trained the network to recognize injured tissue, which then generalized to human COVID-19 patients.

Motivation for polymorphic training: combining specific labels from healthy-density lungs with nonspecific labels from high-density injured lungs.

Instead of discarding datasets with "imperfect" or generic labels, polymorphic training treats label specificity as a hierarchy. This allows researchers to utilize much larger, more diverse pools of data. Furthermore, by feeding the resulting segmentations into a lobe-segmentation algorithm (LobeNet), the authors identified four distinct "radiographical phenotypes" of COVID-19 based on the distribution of consolidation across different lung lobes.

Datasets and Evaluation

The study utilized a diverse training set consisting of 1,000 human CTs from the COPDGene study and 453 animal CTs (canine, porcine, ovine) with induced acute lung injury. Evaluation was conducted on nearly 8,000 CT scans across four pathologies:

  • COPDGene: 5,986 scans.
  • Lung Cancer: 1,620 scans.
  • Idiopathic Pulmonary Fibrosis (IPF): 305 scans.
  • COVID-19: 87 scans for quantitative validation and 133 for phenotype analysis.

Key Results

  • High Accuracy on COVID-19: Despite the absence of COVID-19 data in the training set, the model achieved a Dice coefficient of 0.985 and an average symmetric surface distance (ASSD) of 0.495 mm on COVID-19 cases.
  • Robustness to Consolidation: While standard models trained only on human COPD data failed as lung opacification increased, the polymorphic model maintained high performance even in severe cases.
  • Efficiency: The automated pipeline processes a scan in approximately 2.5 minutes on a GPU, a significant improvement over the several hours required for manual expert segmentation.

Limitations and Future Work

The authors note that at the time of the study, there was a lack of other specialized automated tools for COVID-19 to serve as a direct comparison. Additionally, the regional lobar analysis was performed on a relatively small cohort of 133 cases. Future research may extend polymorphic training directly to lobar segmentation, enabling networks to learn lobe boundaries even when they are obscured by severe disease.

Summary updated

BibTeX
@misc{gerard2020ct,
  author = {Sarah E. Gerard and Jacob Herrmann and Yi Xin and Kevin T. Martin and Emanuele Rezoagli and Davide Ippolito and Giacomo Bellani and Maurizio Cereda and Junfeng Guo and Eric A. Hoffman and David W. Kaczka and Joseph M. Reinhardt},
  title = {{CT Image Segmentation for Inflamed and Fibrotic Lungs Using a Multi-Resolution Convolutional Neural Network}},
  year = {2020},
  eprint = {2010.08582},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.08582},
}
Figure