Source-linked AI summary
CovidAID: COVID-19 Detection Using Chest X-Ray
Arpan Mangal, Surya Kalia, Harish Rajgopal, Krithika Rangarajan, Vinay Namboodiri, Subhashis Banerjee, Chetan Arora
TL;DR
Limited testing resources motivate using chest X-Rays to triage symptomatic patients for further COVID-19 testing. CovidAID adapts a deep neural network for automated classification and reports 90.5% accuracy with 100% COVID-19 sensitivity, while acknowledging the small public dataset.
Problem
Limited testing kits, long turnaround times, and limited sensitivity motivate chest X-Ray-based triage for patients with respiratory illness awaiting COVID-19 testing.
Method
CovidAID uses a deep neural network to classify chest X-Rays into normal, bacterial pneumonia, viral pneumonia, and COVID-19 classes for testing triage.
Results
90.5% accuracy was obtained for the three-class configuration, with 1.0 sensitivity for the COVID-19 positive class.
Takeaways & Limitations
The model is presented as a triage tool to prioritize patients for COVID-19 testing, not as an alternative to conventional diagnostic tests.
Takeaways & Limitations
The reported results are initial and based on a small publicly available dataset requiring further validation with larger datasets and clinical trials.
Abstract
from arXiv · showhide
The exponential increase in COVID-19 patients is overwhelming healthcare systems across the world. With limited testing kits, it is impossible for every patient with respiratory illness to be tested using conventional techniques (RT-PCR). The tests also have long turn-around time, and limited sensitivity. Detecting possible COVID-19 infections on Chest X-Ray may help quarantine high risk patients while test results are awaited. X-Ray machines are already available in most healthcare systems, and with most modern X-Ray systems already digitized, there is no transportation time involved for the samples either. In this work we propose the use of chest X-Ray to prioritize the selection of patients for further RT-PCR testing. This may be useful in an inpatient setting where the present systems are struggling to decide whether to keep the patient in the ward along with other patients or isolate them in COVID-19 areas. It would also help in identifying patients with high likelihood of COVID with a false negative RT-PCR who would need repeat testing. Further, we propose the use of modern AI techniques to detect the COVID-19 patients using X-Ray images in an automated manner, particularly in settings where radiologists are not available, and help make the proposed testing technology scalable. We present CovidAID: COVID-19 AI Detector, a novel deep neural network based model to triage patients for appropriate testing. On the publicly available covid-chestxray-dataset [2], our model gives 90.5% accuracy with 100% sensitivity (recall) for the COVID-19 infection. We significantly improve upon the results of Covid-Net [10] on the same dataset.
1 Introduction
The paper proposes automated chest X-Ray analysis to help triage patients with severe respiratory illness for COVID-19 testing. CovidAID is intended as a testing-support tool rather than a replacement for conventional diagnosis.
- Limited diagnostic kits, hospital beds, PPE, and ventilators motivate distinguishing SARI patients who may have COVID-19.
- CovidAID classifies frontal-view X-Rays as normal, bacterial pneumonia, viral pneumonia, or COVID-19 pneumonia.
- Chest X-Ray is presented as widespread, cost effective, rapidly transferable digitally, and usable with portable machines inside isolation wards.
- The proposed deep neural network is designed to support analysis when radiologists or specialized COVID-19 expertise are unavailable.
- The model is explicitly intended to triage patients for COVID-19 testing, not replace conventional diagnostic tests.
- The authors caution that the model and report reflect current understanding based on limited data and may be updated.
2 Related Work
The work builds on deep-learning approaches for thoracic disease and positions COVID-Net as an open-source baseline for COVID-19 X-Ray classification. Public COVID-19 X-Ray applications and datasets remain limited.
- CheXNet uses a 121-layer DenseNet trained on 112,120 frontal-view X-Rays to classify 14 thoracic disease classes, including pneumonia.
- The authors select CheXNet as a backbone because its input samples and pneumonia task are visually close to COVID-19 pneumonia detection.
- Open-source chest X-Ray screening applications and publicly available COVID-19 X-Ray data are limited.
- COVID-Net is described as the only actively maintained open-source tool among the cited approaches that identifies COVID-19 and other pneumonia classes.
- COVID-Net takes a chest X-Ray as input and predicts Normal, Pneumonia, or COVID-19, serving as the comparison baseline.
3 Method
CovidAID adapts a pretrained CheXNet DenseNet for multi-class chest X-Ray classification, using class-wise binary outputs and weighted loss. It evaluates both four-class and pneumonia-clubbed three-class configurations.
- 3.1 Problem Formulation and Loss Function: The four-class configuration predicts Normal, Bacterial Pneumonia, Viral Pneumonia, and COVID-19, while the three-class configuration combines bacterial and viral pneumonia.
- 3.1 Problem Formulation and Loss Function: Each class is treated as a binary classification problem, with a frontal-view X-Ray as input and class-presence labels as outputs.
- 3.1 Problem Formulation and Loss Function: The model uses weighted binary cross-entropy, weighting positive and negative terms according to class sample counts.
- 3.3 Training: CovidAID uses a pretrained 121-layer CheXNet DenseNet with a replacement classifier producing four or three class outputs.
- 3.3 Training: Training first freezes the DenseNet backbone while fitting the classifier, then fine-tunes the whole network end-to-end.
4 Dataset and Evaluation
The evaluation combines COVID-19 X-Rays with bacterial, viral, and normal chest X-Rays and uses patient-aware splitting to limit information leakage. Training batches subsample pneumonia data because COVID-19 examples are scarce.
- The combined datasets contain COVID-19, bacterial pneumonia, viral pneumonia, and normal frontal-view chest X-Ray images, with CheXNet supplying pretrained features.
- The COVID-19 dataset is split by patient ID to prevent information leakage from multiple images of the same patient.
- The evaluation uses 20% of images as a test set and retains 10 images for validation.
- Because COVID-19 data are scarce, each training batch uses a random subset of pneumonia data to prevent non-COVID losses from masking COVID-19 loss.
5 Results
CovidAID achieves strong discrimination and classification performance across three- and four-class chest X-Ray configurations, including perfect COVID-19 sensitivity, and outperforms COVID-Net on the reported comparison.
- 0.9994 AUROC was achieved for the COVID-19 positive class, with a mean AUROC of 0.9738 for four-class classification.
- 90.5% accuracy was obtained for three-class classification, compared with 87.2% for four-class classification.
- Viral pneumonia reached 0.87 sensitivity and was often confused with bacterial pneumonia, likely because of overlapping imaging characteristics.
- 1.0 sensitivity was achieved for the COVID-19 positive class, matching the sensitivity reported for bacterial pneumonia.
- 5.1 Comparison with COVID-Net: >0.14 AUROC improvement over COVID-Net was reported for detecting both regular pneumonia and COVID-19 on the evaluated test split.The comparison used COVID-Net’s Small variant; the COVID-19 data were the same, while the pneumonia dataset differed.
- 5.1 Comparison with COVID-Net: CovidAID showed higher F1 scores than COVID-Net across 10,000 bootstrap samples from a 654-image test set.The comparison included 95% confidence intervals, and the authors describe the performance difference as substantial.
6 Conclusion
The study presents initial evidence that a deep-learning model can detect COVID-19 from chest X-rays and reports improved performance over COVID-Net. However, the findings remain preliminary because the publicly available dataset is small, motivating validation on larger datasets and in clinical trials.
- The model showed significant performance improvement over COVID-Net for COVID-19 chest X-ray classification.The comparison used the same chest-xray-pneumonia dataset.
- The reported results are promising but preliminary because the publicly available COVID-19 X-ray dataset is small.
- The authors plan further validation using larger COVID-19 X-ray datasets and clinical trials.