Towards computer-aided severity assessment via deep neural networks for geographic and opacity extent scoring of SARS-CoV-2 chest X-rays

Alexander Wong, Zijin Lin, Lanying Wang, Audrey G. Chung, Beiyi Shen, Almas Abbasi, Mahsa Hoshmand-Kochi, Timothy Q. DuongScientific 2021 · DOI

Abstract

A critical step in effective care and treatment planning for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the cause for the coronavirus disease 2019 (COVID-19) pandemic, is the assessment of the severity of disease progression. Chest x-rays (CXRs) are often used to assess SARS-CoV-2 severity, with two important assessment metrics being extent of lung involvement and degree of opacity. In this proof-of-concept study, we assess the feasibility of computer-aided scoring of CXRs of SARS-CoV-2 lung disease severity using a deep learning system. Data consisted of 396 CXRs from SARS-CoV-2 positive patient cases. Geographic extent and opacity extent were scored by two board-certified expert chest radiologists and a 2nd-year radiology resident. The deep neural networks used in this study, which we name COVID-Net S, are based on a COVID-Net network architecture. The COVID-Net S deep neural networks yielded $R^2$ of 0.664 and 0.635 between predicted scores and radiologist scores for geographic extent and opacity extent, respectively. The results suggest that deep neural networks could be an effective tool for computer-aided assessment of SARS-CoV-2 lung disease severity.

Introduction

During the COVID-19 pandemic, assessing disease progression has been critical for clinical decision-making, such as determining the need for intensive care or mechanical ventilation. While Computed Tomography (CT) scans provide high detail, they are expensive and pose logistical challenges for infectious patients. Chest X-rays (CXRs) offer a faster, more portable alternative, but manual interpretation for severity is subjective and time-consuming. This study introduces COVID-Net S, a deep learning framework designed to standardize and automate the quantification of lung involvement via geographic and opacity extent scores.

Methods and Architecture

The researchers developed COVID-Net S by adapting the COVID-Net backbone for regression tasks. Unlike standard classification models, COVID-Net S predicts continuous severity scores.

The COVID-Net S architecture, highlighting the PEPX blocks and the regression-based output layers for severity scoring.

Architectural Design

The network utilizes PEPX (Projection-Expansion-Projection) design patterns. These blocks provide high representational capacity and architectural diversity while remaining computationally efficient. The final layers of the original COVID-Net were replaced with a 128-neuron dense layer, a 3-neuron dense layer, and a single output layer to predict a severity score on a continuous scale (0 to 1), which is then mapped back to the 0–8 clinical scale used by radiologists.

Training and Validation

The model was pre-trained on the COVIDx dataset (nearly 14,000 images) to leverage features from general pneumonia detection. To ensure robustness, the team employed stratified Monte Carlo cross-validation, running 100 independent trials (50 for geographic extent and 50 for opacity) to evaluate performance across different data subsets.

Dataset and Expert Scoring

The study utilized a dataset of 396 CXRs from 267 COVID-19 positive patients. The images represented a diverse global pool with various imaging views (AP/PA) and patient positions.

Representative chest X-rays illustrating the spectrum of geographic extent severity used for model training and evaluation.

Three experts—two board-certified radiologists with over 20 years of experience and one radiology resident—provided the ground truth. Each lung was scored 0–4 for geographic extent and 0–4 for opacity, resulting in a total possible score of 8 per metric. High inter-reader agreement (ICC of 0.87–0.92) ensured a reliable gold standard for training.

Key Results

The performance was measured using the coefficient of determination ($R^2$) between the AI's predicted scores and the radiologists' average scores:

  • Geographic Extent: Achieved a mean $R^2$ of 0.664 ± 0.032, with the best-performing individual network reaching 0.739.
  • Opacity Extent: Achieved a mean $R^2$ of 0.635 ± 0.044, with the best-performing network reaching 0.741.

The study observed a strong correlation between geographic extent and opacity, suggesting that spatial spread and inflammatory density typically increase in tandem. Furthermore, the PEPX architecture allowed the model to maintain "selective long-range connectivity," improving learning efficiency for deployment in resource-constrained settings.

Discussion and Future Work

While these results are promising, the study notes several areas for further development. The current severity scores are based on radiologist perception rather than direct patient outcomes (e.g., mortality or ventilation requirements). Future research should correlate AI scores with longitudinal clinical data to track how a patient's condition changes over time. Additionally, expanding the dataset to include more diverse imaging equipment will help mitigate potential hardware-related biases. COVID-Net S represents a significant step toward automated clinical triaging and standardized severity reporting for COVID-19.

Summary updated

Abstract

A critical step in effective care and treatment planning for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the cause for the coronavirus disease 2019 (COVID-19) pandemic, is the assessment of the severity of disease progression. Chest x-rays (CXRs) are often used to assess SARS-CoV-2 severity, with two important assessment metrics being extent of lung involvement and degree of opacity. In this proof-of-concept study, we assess the feasibility of computer-aided scoring of CXRs of SARS-CoV-2 lung disease severity using a deep learning system. Data consisted of 396 CXRs from SARS-CoV-2 positive patient cases. Geographic extent and opacity extent were scored by two board-certified expert chest radiologists (with 20+ years of experience) and a 2nd-year radiology resident. The deep neural networks used in this study, which we name COVID-Net S, are based on a COVID-Net network architecture. 100 versions of the network were independently learned (50 to perform geographic extent scoring and 50 to perform opacity extent scoring) using random subsets of CXRs from the study, and we evaluated the networks using stratified Monte Carlo cross-validation experiments. The COVID-Net S deep neural networks yielded R[Formula: see text] of [Formula: see text] and [Formula: see text] between predicted scores and radiologist scores for geographic extent and opacity extent, respectively, in stratified Monte Carlo cross-validation experiments. The best performing COVID-Net S networks achieved R[Formula: see text] of 0.739 and 0.741 between predicted scores and radiologist scores for geographic extent and opacity extent, respectively. The results are promising and suggest that the use of deep neural networks on CXRs could be an effective tool for computer-aided assessment of SARS-CoV-2 lung disease severity, although additional studies are needed before adoption for routine clinical use.
BibTeX
@inproceedings{wong2021towards,
  author = {Alexander Wong and Zijin Lin and Lanying Wang and Audrey G. Chung and Beiyi Shen and Almas Abbasi and Mahsa Hoshmand-Kochi and Timothy Q. Duong},
  title = {{Towards computer-aided severity assessment via deep neural networks for geographic and opacity extent scoring of SARS-CoV-2 chest X-rays}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-88538-4},
}
Figure