Source-linked AI summary
COVIDNet-CT: A Tailored Deep Convolutional Neural Network Design for Detection of COVID-19 Cases from Chest CT Images
Hayden Gunraj, Linda Wang, Alexander Wong
TL;DR
COVID-19 screening requires alternatives or complements to RT-PCR because sensitivity varies and results can be delayed, while CT abnormalities may overlap subtly with other pneumonia. The paper develops COVIDNet-CT using machine-driven design exploration, introduces the COVIDx-CT dataset, and validates predictions through explainability analysis. On COVIDx-CT, COVIDNet-CT achieves 99.1% test accuracy, while the authors report that its decisions leverage lung abnormalities; the model is not yet suitable for clinical use.
Problem
COVID-19 screening needs fast and effective tools because RT-PCR sensitivity varies with sampling and symptom timing, results may be delayed, and CT abnormalities can resemble other pneumonia.
Method
The study develops COVIDNet-CT through machine-driven design exploration, constructs COVIDx-CT from CNCB data, and uses explainability-driven validation to audit prediction features.
Results
99.1% test accuracy was achieved by COVIDNet-CT on the COVIDx-CT dataset, and explainability analysis found that lung abnormalities were leveraged in decision-making.
Takeaways & Limitations
COVIDNet-CT and COVIDx-CT were released openly to encourage research and improvement, while explainability analysis may aid understanding of CT-based screening.
Takeaways & Limitations
COVIDNet-CT is not yet suitable for clinical use, and improving generalizability may require expanding and diversifying COVIDx-CT.
Abstract
from arXiv · showhide
The coronavirus disease 2019 (COVID-19) pandemic continues to have a tremendous impact on patients and healthcare systems around the world. In the fight against this novel disease, there is a pressing need for rapid and effective screening tools to identify patients infected with COVID-19, and to this end CT imaging has been proposed as one of the key screening methods which may be used as a complement to RT-PCR testing, particularly in situations where patients undergo routine CT scans for non-COVID-19 related reasons, patients with worsening respiratory status or developing complications that require expedited care, and patients suspected to be COVID-19-positive but have negative RT-PCR test results. Motivated by this, in this study we introduce COVIDNet-CT, a deep convolutional neural network architecture that is tailored for detection of COVID-19 cases from chest CT images via a machine-driven design exploration approach. Additionally, we introduce COVIDx-CT, a benchmark CT image dataset derived from CT imaging data collected by the China National Center for Bioinformation comprising 104,009 images across 1,489 patient cases. Furthermore, in the interest of reliability and transparency, we leverage an explainability-driven performance validation strategy to investigate the decision-making behaviour of COVIDNet-CT, and in doing so ensure that COVIDNet-CT makes predictions based on relevant indicators in CT images. Both COVIDNet-CT and the COVIDx-CT dataset are available to the general public in an open-source and open access manner as part of the COVID-Net initiative. While COVIDNet-CT is not yet a production-ready screening solution, we hope that releasing the model and dataset will encourage researchers, clinicians, and citizen data scientists alike to leverage and build upon them.
1 Introduction
COVID-19 screening needs are heightened by RT-PCR sensitivity variability and delays, while CT abnormalities can be subtle and overlap with non-COVID-19 pneumonia. The paper introduces COVIDNet-CT, COVIDx-CT, and explainability-based validation to address this challenge.
- Screening motivation: RT-PCR is highly specific but has variable sensitivity and can delay screening results.Sensitivity depends on sampling method and time since symptom onset.
- Screening motivation: CT imaging may complement RT-PCR, especially during routine scans, worsening respiratory status, or suspected infection despite negative RT-PCR.Routine examinations can be leveraged without additional COVID-19-specific CT costs or resource constraints.
- Imaging challenge: COVID-19-associated CT abnormalities include ground-glass opacities, patchy shadows, crazy-paving patterns, and consolidation, often bilaterally.Such abnormalities may appear despite a negative RT-PCR test.
- Imaging challenge: Visual differences between COVID-19 and non-COVID-19 pneumonia abnormalities can be subtle, complicating radiologist discrimination.Figure 1 illustrates the subtle distinction between representative cases.
- Study contribution: The study introduces COVIDNet-CT, COVIDx-CT, and explainability-driven validation, releasing the model and dataset openly through the COVID-Net initiative.The validation examines whether predictions rely on relevant CT image features.
2 Related work
Prior studies used deep learning and explainability for CT-based COVID-19 detection, but the authors position COVIDNet-CT as distinct in its machine-driven architecture design and explainability-driven validation.
- Deep learning systems: Prior CT-based deep learning systems distinguish COVID-19 from normal, non-COVID-19 pneumonia, or other non-COVID-19 cases.Some systems require lung or lesion segmentation, adding a segmentation stage or manual radiologist work.
- Deep learning systems: COVIDNet-CT is described as the first COVID-19 CT detector built using machine-driven design exploration.The claimed distinction concerns architecture design for the task and data at hand.
- Explainability methods: Earlier explainability studies used Grad-CAM, upsampled infection-probability heatmaps, or correlations between clinical parameters and segmented lesions.These approaches investigated relationships between imaging features and network predictions.
- Explainability methods: The paper describes its validation as the first CT COVID-19 study using explainability to identify specific critical factors rather than only general image-importance heatmaps.The stated purpose is to audit whether network decisions use relevant image features.
3 Methods
The methods construct the COVIDx-CT benchmark, design COVIDNet-CT through constrained machine-driven exploration, and validate its decision-making with explainability analysis.
- COVIDx-CT dataset: COVIDx-CT contains 104,009 chest CT images from 1,489 patient cases, divided approximately 60%-20%-20% into training, validation, and test sets.The dataset uses abnormal-lung slices for NCP and CP volumes and excludes volumes with background-removed segmented lung regions to reduce model bias.
- Machine-driven design exploration: Machine-driven design exploration formulates architecture selection as constrained optimization using a universal performance function and task-specific operational requirements.The validation constraints require COVID-19 sensitivity ≥95% and COVID-19 PPV ≥95%.
- Network architecture: COVIDNet-CT combines heterogeneous spatial, pointwise, and depthwise convolutions with selective long-range connectivity to balance accuracy and computational complexity.The architecture was produced through machine-driven exploration and exhibits substantial microarchitecture variation across layers.
- Network architecture: PRPE and PRPE-S patterns use pointwise projection, channel replication, depthwise spatial processing, projection, and channel expansion to provide computational efficiency with representational capacity.PRPE is unstrided, whereas PRPE-S uses strided depthwise convolutions.
- Implementation details: The model is ImageNet-pretrained, trained on COVIDx-CT with stochastic gradient descent and momentum, and augmented through geometric and intensity transformations.Augmentations include cropping-box jitter, rotation, shearing, flipping, and intensity shifting and scaling.
- Explainability-driven validation: Explainability-driven performance analysis audits which CT imaging features influence COVIDNet-CT predictions because scalar metrics cannot explain network decision-making.The audit is motivated by the clinical importance of understanding decisions that may affect patient health.
4 Experimental results
COVIDNet-CT was evaluated quantitatively and qualitatively on COVIDx-CT, achieving high accuracy and efficiency while showing diagnostically relevant decision-making. It also outperformed ResNet-50 across reported performance and complexity measures.
- Quantitative results: COVIDNet-CT achieved 99.1% test accuracy on COVIDx-CT while maintaining relatively low architectural and computational complexity.The evaluation reported image-level accuracy, sensitivity, and PPV, with Table 1 covering parameters and FLOPs.
- Quantitative results: The authors attribute the efficient, accurate architecture to machine-driven design exploration tailored to the task, data, and operational requirements.They identify computationally limited and embedded clinical scenarios as potential use contexts.
- Quantitative results: COVIDNet-CT achieved 97.3% COVID-19 sensitivity, corresponding to a low proportion of COVID-19 cases misclassified as non-COVID-19 or normal.The authors also report high COVID-19 PPV and fewer non-COVID-19 cases misclassified as COVID-19.
- Architecture comparison: COVIDNet-CT exceeded ResNet-50 accuracy by 0.4% while using 94.1% fewer parameters and 90.2% fewer FLOPs.It also achieved higher sensitivity and specificity than ResNet-50 across all infection types.
- Qualitative results: Explainability-driven validation indicated that COVIDNet-CT primarily used lung abnormalities to identify COVID-19 and distinguish it from non-COVID-19 pneumonia.The analysis was intended to assess whether predictions relied on diagnostically relevant rather than irrelevant visual indicators.
5 Conclusion
The study introduced COVIDNet-CT and COVIDx-CT, evaluated the model quantitatively, and examined its prediction behaviour using explainability analysis. The authors report strong test performance and lung-abnormality-based decisions, while noting that the model is not yet suitable for clinical use.
- COVIDNet-CT is a tailored deep convolutional neural network for chest CT-based COVID-19 detection, evaluated on the COVIDx-CT test dataset.The evaluation considered accuracy, sensitivity, and PPV.
- COVIDx-CT contains 104,009 chest CT images from 1,489 patients.
- Explainability-driven validation found that COVIDNet-CT leveraged abnormalities in the lungs when making predictions.The analysis was intended to assess whether predictions were based on relevant image features.
- COVIDNet-CT and instructions for constructing COVIDx-CT were publicly released to support research-community usage and improvement.
- COVIDNet-CT is not yet suitable for clinical use, and future performance and generalizability may depend on expanding and diversifying COVIDx-CT.The paper also identifies additional clinical tasks as possible future extensions, contingent on high-quality datasets.