Automated Quantification of CT Patterns Associated with COVID-19 from Chest CT.

Shikha Chaganti, Abishek Balachandran, Guillaume Chabin, Stuart L. Cohen, Thomas Flohr, Bogdan Georgescu, Philippe Greniér, Saša Grbić, Siqi Liu, François Mellot, Nicolas Murray, Savvas Nicolaou, William Parker, Thomas J. Re, Pina C. Sanelli, Alexander Sauter, Zhoubing Xu, Youngjin Yoo, Valentin Ziebandt, Dorin ComaniciuPubMed 2020

TL;DR

This paper introduces an automated deep learning system that quantifies COVID-19 related abnormalities in chest CT scans, providing objective measures of disease severity. The system achieves high correlation with expert annotations, offering a faster and more reproducible alternative to manual assessment.

What Problem Does This Paper Address?

Diagnosing and assessing the severity of COVID-19 using chest CT scans is crucial, especially when RT-PCR tests are inconclusive or unavailable. However, manual quantification of lung abnormalities like ground glass opacities and consolidations is time-consuming and prone to subjective variability among radiologists. This leads to challenges in consistent severity assessment, patient triage, and tracking disease progression. The paper aims to bridge this gap by developing an automated system that can accurately and efficiently quantify these abnormalities.

Methods

The proposed system employs a deep learning approach for automated quantification. It involves several key steps:

  • Lung and Lobe Segmentation: A multi-scale deep reinforcement learning algorithm is used to robustly detect anatomical landmarks and segment the lungs and their lobes in 3D. This is followed by a Deep Image-to-Image Network (DI2IN) for precise lung segmentation.
  • Abnormality Segmentation: A DenseUNet architecture is utilized for semantic segmentation, identifying voxels within the lungs that contain ground glass opacities or consolidations. This is treated as a binary classification problem.
  • Severity Quantification: Two combined severity measures are computed:
  • (PO, PHO): Global measures representing the overall percentage of lung opacity (PO) and the percentage of high opacity (PHO), which corresponds to consolidation.
  • (LSS, LHOS): Lobe-wise measures quantifying the lung severity score (LSS) and lung high opacity score (LHOS) for each of the five lung lobes. These scores are derived from the extent of involvement and high opacity within each lobe.

The system is trained using a combination of COVID-19, viral pneumonia, and other interstitial lung disease (ILD) datasets to enhance robustness.

Datasets

The study utilized several datasets:

  • Training Data: A total of 9749 chest CT volumes were used for training and model selection. This included:
  • 901 volumes for abnormality segmentation, comprising 431 COVID-19, 174 viral pneumonia, and 296 ILD datasets.
  • 8792 volumes for lung and lobe segmentation.
  • Testing Data: A separate test set of 200 volumes was used for evaluation, consisting of 100 COVID-19 positive cases and 100 healthy controls. These datasets were collected from multiple institutions across the USA, Canada, and Europe.

The ground truth for all annotations was established by expert medical annotation engineers under the supervision of board-certified radiologists.

Key Insights and Ideas

  • Automated Quantification is Feasible and Accurate: The study demonstrates that deep learning can effectively automate the quantification of COVID-19 related lung abnormalities, achieving high correlation with expert annotations. This suggests a significant potential for clinical adoption.
  • Dual Severity Measures Capture Different Aspects: The introduction of both global (PO, PHO) and lobe-wise (LSS, LHOS) severity measures provides a comprehensive assessment. This allows for understanding both the overall disease burden and its distribution across different lung regions.
  • High Opacity as a Marker for Severity: The focus on "high opacity" (consolidation) as a key indicator of severe symptoms aligns with existing clinical observations, making the PHO and LHOS metrics particularly valuable.
  • Efficiency Gains: The automated system achieves a processing time of approximately 10 seconds per case, a dramatic improvement over the 30 minutes required for manual annotations. This efficiency is critical in high-volume scenarios like an outbreak.

Key Results

  • High Correlation with Ground Truth: The system achieved excellent Pearson correlation coefficients with ground truth annotations for COVID-19 cases:
  • PO: 0.92 (P < .001)
    • PHO: 0.97 (P < .001)
    • LSS: 0.91 (P < .001)
    • LHOS: 0.90 (P < .001)
  • Accurate Classification of Controls: 98 out of 100 healthy controls were correctly classified with a predicted PO of less than 1%, indicating a low rate of false positives.
  • Rapid Processing: The automated processing time was approximately 10 seconds per case, compared to 30 minutes for manual annotations.
  • Qualitative Validation: Visualizations of segmentation results showed good agreement between the system's output and ground truth for successful cases, with some outliers identified and analyzed.

Limitations and Future Work

  • Limitations:
    • The system was trained on specific abnormalities (GGO and consolidation) relevant to COVID-19. The presence of other abnormalities like pleural effusion could pose challenges.
  • The study focused on quantifying abnormalities in lung CT scans only.
  • The paper acknowledges that the AI system's output values are slightly higher than ground truth, suggesting a need for further calibration for more accurate metric measurements.
  • Future Work:
    • Training the system to account for a broader range of abnormalities beyond GGO and consolidation.
  • Improving the algorithm's performance in boundary conditions to address under-segmentation of faint regions and over-segmentation of highly abnormal regions.
  • Evaluating the system's ability to distinguish COVID-19 from other lung diseases like viral pneumonias and ILDs.
  • Correlating the computed imaging findings with clinical factors and patient outcomes.
  • Exploring the application of this methodology to X-ray chest images.

Summary updated

Abstract

A new method segments regions of CT abnormalities associated with COVID-19 and computes (PO, PHO), as well as (LSS, LHOS) severity scores.
BibTeX
@inproceedings{chaganti2020automated,
  author = {Shikha Chaganti and Abishek Balachandran and Guillaume Chabin and Stuart L. Cohen and Thomas Flohr and Bogdan Georgescu and Philippe Greniér and Saša Grbić and Siqi Liu and François Mellot and Nicolas Murray and Savvas Nicolaou and William Parker and Thomas J. Re and Pina C. Sanelli and Alexander Sauter and Zhoubing Xu and Youngjin Yoo and Valentin Ziebandt and Dorin Comaniciu},
  title = {{Automated Quantification of CT Patterns Associated with COVID-19 from Chest CT.}},
  booktitle = {PubMed},
  year = {2020},
}
Figure