PaperlayerView exact revision on arXiv

Source-linked AI summary

COVID-Net: A Tailored Deep Convolutional Neural Network Design for Detection of COVID-19 Cases from Chest X-Ray Images

Linda Wang, Alexander Wong

arXiv:2003.09871v4eess.IVcs.CVcs.LG

TL;DR

COVID-19 screening needs accessible radiographic approaches. This paper introduces the open-source COVID-Net model, COVIDx dataset, and explainability audit, achieving 93.3% test accuracy and 98.9% COVID-19 PPV on COVIDx while improving transparency.

  • Problem

    The paper addresses the need for effective COVID-19 screening using accessible chest radiography alongside RT-PCR testing.

  • Method

    The study develops open-source COVID-Net, constructs the COVIDx benchmark dataset, and audits model decisions using explainability analysis.

  • Results

    93.3% test accuracy and 98.9% COVID-19 PPV were achieved on COVIDx, with higher test accuracy and COVID-19 sensitivity than VGG-19 and ResNet-50.

  • Takeaways & Limitations

    Explainability analysis is intended to improve trust and transparency when leveraging COVID-Net for computer-aided screening.

  • Takeaways & Limitations

    COVID-Net is not production-ready, and the authors identify sensitivity and COVID-19 PPV as areas for improvement with additional data.

Abstract

from arXiv · show

The COVID-19 pandemic continues to have a devastating effect on the health and well-being of the global population. A critical step in the fight against COVID-19 is effective screening of infected patients, with one of the key screening approaches being radiology examination using chest radiography. Motivated by this and inspired by the open source efforts of the research community, in this study we introduce COVID-Net, a deep convolutional neural network design tailored for the detection of COVID-19 cases from chest X-ray (CXR) images that is open source and available to the general public. To the best of the authors' knowledge, COVID-Net is one of the first open source network designs for COVID-19 detection from CXR images at the time of initial release. We also introduce COVIDx, an open access benchmark dataset that we generated comprising of 13,975 CXR images across 13,870 patient patient cases, with the largest number of publicly available COVID-19 positive cases to the best of the authors' knowledge. Furthermore, we investigate how COVID-Net makes predictions using an explainability method in an attempt to not only gain deeper insights into critical factors associated with COVID cases, which can aid clinicians in improved screening, but also audit COVID-Net in a responsible and transparent manner to validate that it is making decisions based on relevant information from the CXR images. By no means a production-ready solution, the hope is that the open access COVID-Net, along with the description on constructing the open source COVIDx dataset, will be leveraged and build upon by both researchers and citizen data scientists alike to accelerate the development of highly accurate yet practical deep learning solutions for detecting COVID-19 cases and accelerate treatment of those who need it the most.

1 Introduction

The introduction frames rapid, accessible chest X-ray screening as a useful complement to RT-PCR during the COVID-19 pandemic, while identifying expert interpretation as a bottleneck. It presents COVID-Net, an open-source deep convolutional neural network designed to help detect COVID-19 cases from CXR images.

  • Effective COVID-19 screening is critical for enabling immediate treatment, care, and isolation of infected patients.
  • CXR imaging supports rapid triaging and is widely available, more accessible than CT, and deployable through portable systems.These properties are especially valuable in resource-constrained and heavily affected areas.
  • Expert radiologist interpretation is a major bottleneck because COVID-19 visual indicators in radiography images can be subtle.The introduction therefore motivates computer-aided diagnostic systems that can support faster and more accurate interpretation.
  • The study introduces COVID-Net, an open-source deep convolutional neural network tailored to detect COVID-19 cases from CXR images.The design is motivated by the urgent pandemic response and open-source efforts seeking to leverage accessible CXR imaging.

2 Related Work

Earlier COVID-19 radiography AI research showed promising detection accuracy but focused primarily on CT imaging. After COVIDx and COVID-Net were released, subsequent CXR studies often used their variants and predominantly explored deep convolutional neural networks.

  • Prior radiography AI systems: Earlier deep-learning AI systems reported promising accuracy for detecting COVID-19 through radiography, with research focused primarily on CT imaging.These systems were motivated by the need for faster interpretation of radiography images.
  • Subsequent CXR studies: Subsequent COVID-19 CXR studies often leveraged variants of the COVIDx dataset or COVID-Net.The passage identifies multiple studies conducted after the initial public releases of COVIDx and COVID-Net.
  • Subsequent CXR studies: The majority of subsequent studies explored deep convolutional neural networks, reflecting their successes across computer vision tasks.The passage describes these successes as significant and state-of-the-art achievements in a variety of computer vision applications.

3 Methods

The methods use a human-machine collaborative strategy to develop COVID-Net, combining principled human network design with machine-driven exploration for COVID-19 detection from chest X-ray images. The study also creates the open-access COVIDx benchmark dataset for training and evaluating COVID-Net.

  • COVID-Net design: COVID-Net is developed through human-driven principled network design prototyping combined with machine-driven design exploration.This collaborative process produces an architecture tailored for detecting COVID-19 cases from chest X-ray images.
  • COVIDx dataset: The study creates COVIDx, an open-access benchmark dataset to facilitate training and evaluating COVID-Net.

3.1 COVIDx Dataset

COVIDx is an open-access benchmark dataset containing 13,975 chest X-ray images from 13,870 patient cases, assembled by combining and modifying five public repositories. Its limited COVID-19 case data reflects public-domain scarcity and motivates continued dataset expansion.

  • Dataset overview: 13,975 CXR images across 13,870 patient cases comprise the COVIDx dataset, described as the largest open-access benchmark by COVID-19-positive patient cases.The dataset was used to train and evaluate COVID-Net.
  • Dataset construction: Five publicly available repositories were combined and modified to generate COVIDx.The sources included COVID-19 collections, the RSNA Pneumonia Detection Challenge dataset, and the COVID-19 radiography database.
  • Case composition: COVIDx includes COVID-19, non-COVID-19 pneumonia, and normal patient cases drawn from the contributing repositories.COVID-19 cases came from multiple repositories, while RSNA contributed normal and non-COVID-19 pneumonia cases.
  • Dataset limitations: Limited COVID-19 infection cases and associated CXR images reflect scarce public-domain data and underscore the need for additional cases to improve COVIDx.The dataset contains 358 CXR images from 266 COVID-1… cases in the supplied passage.

3.2 Principled network design prototyping

The initial COVID-Net prototype was designed using human-driven principles and residual architecture practices. It was structured to predict normal, non-COVID-19 infection, or COVID-19 infection, supporting clinical prioritization for PCR confirmation.

  • Principled network design prototyping: The initial prototype used human-driven design principles and residual architecture practices to support reliable, trainable high-performance neural networks.Residual architectures were selected because they have repeatedly enabled reliable networks that are easier to train to high performance.
  • Principled network design prototyping: The prototype classified CXRs into three outcomes: no infection, non-COVID-19 infection, or COVID-19 viral infection.The three-way prediction was intended to help clinicians prioritize patients for PCR testing and case confirmation.

3.3 Machine-driven Design Exploration

Machine-driven design exploration uses the initial prototype, data, and human-specified requirements to identify an optimal macroarchitecture and microarchitecture for COVID-Net. The study leverages generative synthesis, using a generator-inquisitor pair to learn architectures satisfying those requirements.

  • Machine-driven Design Exploration: The machine-driven stage uses the initial network prototype, data, and human-specific design requirements to guide exploration of COVID-Net’s final architecture.It targets both macroarchitecture and microarchitecture designs.
  • Machine-driven Design Exploration: COVID-Net combines high architectural diversity, selective long-range connectivity, and projection-expansion-projection patterns to increase representational capacity while maintaining computational efficiency.
  • Machine-driven Design Exploration: Machine-driven exploration provides greater granularity and flexibility than manual human-driven architecture design while retaining human-specified design requirements.
  • Machine-driven Design Exploration: Generative synthesis uses a generator-inquisitor pair to learn and generate deep neural network architectures that satisfy human-specified design requirements.The pair works in tandem to garner insights during design exploration.

3.4 COVID-Net Network Architecture

COVID-Net is an openly available architecture created through human-machine collaborative design, combining a lightweight PEPX pattern, selective long-range connectivity, and heterogeneous convolutional components. Its customized design aims to enhance representational capacity while reducing computational complexity.

  • Architecture design: COVID-Net is publicly available and was created through a human-machine collaborative design strategy.The architecture is available at https://github.com/lindawangg/COVID-Net.
  • PEPX design: COVID-Net heavily uses a lightweight residual projection-expansion-projection-extension (PEPX) pattern.The pattern applies first-stage projection, expansion, depth-wise representation, second-stage projection, and extension operations.
  • PEPX design: The customized PEPX pattern was not previously introduced in literature to the authors’ knowledge and enables enhanced representational capacity with reduced computational complexity.Efficient 3×3 depth-wise convolutions learn spatial characteristics while minimizing computational complexity and preserving representational capacity.
  • Connectivity: COVID-Net incorporates selective long-range connectivity across various areas of the architecture.The passage associates long-range connectivity with improved representational capacity and easier training, while noting disadvantages when many such connections are used.
  • Architectural diversity: COVID-Net has architectural diversity, mixing convolution kernels from 7×7 to 1×1 and grouping configurations from ungrouped to depth-wise.This heterogeneous mixture reinforces the role of machine-driven design exploration in the architecture.

3.5 Implementation Details

COVID-Net was pretrained on ImageNet36 and trained on COVIDx with Adam under a patience-based learning-rate schedule. Training used specified hyperparameters, data augmentation, batch re-balancing, and a Keras/TensorFlow implementation built with generative synthesis.

  • Training setup: COVID-Net was pretrained on ImageNet36, then trained on COVIDx using Adam with a learning-rate schedule that decreases after learning stagnates.The schedule used factor=0.7 and patience=5.
  • Training setup: 2e-4 learning rate, 22 epochs, and batch size 64 were used for training.These hyperparameters were specified for the COVID-Net training procedure.
  • Data processing: Data augmentation included translation, rotation, horizontal flip, zoom, and intensity shift.These transformations were leveraged during training.
  • Data processing: Batch re-balancing was introduced to promote better distribution of each infection type at the batch level.The strategy targeted infection-type distribution within individual batches.
  • Implementation: The initial COVID-Net prototype used Keras with a TensorFlow backend, and its architecture was built using generative synthesis.These tools supported the prototype's construction and evaluation.

3.6 COVID-Net Auditing via Explainability

The study audits COVID-Net with explainability to assess whether its COVID-19 detection decisions rely on relevant CXR information rather than improper visual indicators. It uses GSInquire’s generator–inquisitor framework to derive interpretations of decisions made by the reference network.

  • Audit objective: The explainability-driven audit evaluates whether COVID-Net bases clinical detection decisions on relevant information instead of improper visual indicators.The motivation is responsibility and transparency because COVID-19 detection can affect patients’ health and well-being.
  • GSInquire framework: GSInquire pairs a generator G with an inquisitor I to obtain insights into deep neural networks and learn to generate networks.The inquisitor’s insights can subsequently be transformed into interpretations of network decisions.
  • Decision interpretation: For COVID-Net and a CXR input, GSInquire seeks an interpretation z of the reference network’s decision.The reference network is COVID-Net and the input signal is a chest X-ray image.
  • Inquisitor procedure: The inquisitor probes generated networks with the target CXR stimulus, observes reactionary responses, and updates its parameters using those responses and generator-related functions.This process enables the inquisitor to learn from the insights derived from the observed response signals.

4 Experimental Results

COVID-Net achieved strong COVID-19 detection performance on the COVIDx test dataset, including 93.3% test accuracy, 91.0% sensitivity, and 98.9% PPV. Its lightweight architecture reduced computational complexity, while GSInquire-based auditing supported more transparent decisions and potential clinical insight discovery.

  • Quantitative Results: 93.3% test accuracy was achieved by COVID-Net on the COVIDx test dataset.The evaluation also computed sensitivity and positive predictive value for each infection type.
  • Quantitative Results: 98.9% PPV for COVID-19 cases indicates very few false-positive COVID-19 detections.The passage notes one non-COVID19 patient was misidentified as having COVID-19 viral infection.
  • Architectural Analysis: ∼12× fewer MAC operations than VGG-19 and ∼2.37× fewer MAC operations than ResNet-50 were required by COVID-Net.The lower computational complexity illustrates the benefits of lightweight PEPX design patterns.
  • Explainability Audit: GSInquire identified critical factors in example COVID-19 CXR images to audit whether COVID-Net relied on relevant visual information.The audit was intended to support responsible and transparent design by examining how the network makes detection decisions.
  • Explainability Audit: Understanding COVID-Net’s critical factors can make predictions more transparent for clinicians and potentially reveal visual indicators associated with SARS-CoV-2 infection.These insights could support faster, accurate assessments and improved screening accuracy.

5 Conclusion

The study introduced COVID-Net, an open-source deep convolutional neural network for detecting COVID-19 from chest X-ray images, alongside COVIDx, an open-access benchmark dataset. It also examined COVID-Net’s predictions using explainability methods while emphasizing that the system is not production-ready and may support future research.

  • COVID-Net is an open-source deep convolutional neural network designed to detect COVID-19 cases from chest X-ray images.
  • COVIDx comprises 13,975 chest X-ray images across 13,870 patient cases from five open-access data repositories.
  • The study used an explainability method to investigate how COVID-Net makes predictions and gain insight into factors associated with COVID-19 cases.
  • COVID-Net is explicitly not production-ready, but its promising COVIDx test results and open-source release are intended to encourage further development of practical detection solutions.
Loading 2003.09871v4…