Source-linked AI summary
COVID-ResNet: A Deep Learning Framework for Screening of COVID19 from Radiographs
Muhammad Farooq, Abdul Hafeez
TL;DR
The paper addresses the need for rapid, reliable COVID-19 screening and the limited availability of open datasets and reproducible approaches. It fine-tunes a pretrained ResNet50 using progressive resizing and learning-rate techniques for four-class chest-radiograph classification. COVID-ResNet reaches 96.23% accuracy on COVIDx with 41 epochs, while the authors caution that larger datasets and real-world testing are needed before clinical use.
Problem
COVID-19 screening is needed because existing tests can be slow or limited, while prior imaging approaches and datasets were often not publicly available for reproducible research.
Method
COVID-ResNet fine-tunes an ImageNet-pretrained ResNet50 using progressive image resizing, automatic learning-rate finding, discriminative learning rates, and data augmentation.
Results
96.23% accuracy was achieved on the independent COVIDx test set, compared with 83.5% for COVID-Net, using 25.6M versus 116.6M parameters.
Takeaways & Limitations
The model provides a computationally efficient and accurate four-class classification framework for COVID-19, other pneumonias, and Normal radiographs.
Takeaways & Limitations
COVID-19 test cases were very small compared with other classes, and the authors plan further testing as additional COVID-19 data becomes available.
Abstract
from arXiv · showhide
In the last few months, the novel COVID19 pandemic has spread all over the world. Due to its easy transmission, developing techniques to accurately and easily identify the presence of COVID19 and distinguish it from other forms of flu and pneumonia is crucial. Recent research has shown that the chest Xrays of patients suffering from COVID19 depicts certain abnormalities in the radiography. However, those approaches are closed source and not made available to the research community for re-producibility and gaining deeper insight. The goal of this work is to build open source and open access datasets and present an accurate Convolutional Neural Network framework for differentiating COVID19 cases from other pneumonia cases. Our work utilizes state of the art training techniques including progressive resizing, cyclical learning rate finding and discriminative learning rates to training fast and accurate residual neural networks. Using these techniques, we showed the state of the art results on the open-access COVID-19 dataset. This work presents a 3-step technique to fine-tune a pre-trained ResNet-50 architecture to improve model performance and reduce training time. We call it COVIDResNet. This is achieved through progressively re-sizing of input images to 128x128x3, 224x224x3, and 229x229x3 pixels and fine-tuning the network at each stage. This approach along with the automatic learning rate selection enabled us to achieve the state of the art accuracy of 96.23% (on all the classes) on the COVIDx dataset with only 41 epochs. This work presented a computationally efficient and highly accurate model for multi-class classification of three different infection types from along with Normal individuals. This model can help in the early screening of COVID19 cases and help reduce the burden on healthcare systems.
I. INTRODUCTION
COVID-19 screening is urgently needed because existing tests can be slow, limited in availability, or error-prone, while radiographs may provide useful imaging signals. The paper targets open, accurate classification across COVID-19 and other infection classes using efficient training techniques.
- Motivation: Faster and reliable COVID-19 screening is needed because serology and rRT-PCR can require hours to days and test-kit availability is limited.The paper frames screening as important for quick treatment and patient isolation.
- Motivation: Chest radiography may support initial COVID-19 screening, but radiologists must visually detect subtle abnormalities.The authors motivate automated extraction of radiographic features characteristic of COVID-19.
- Research gap: Publicly unavailable datasets limit broader research access and development of classification techniques on standardized data.The work builds on COVIDx and COVID-Net, whose four-class dataset includes Normal, Bacterial, Viral, and COVID-19 cases.
- Objectives: The study seeks higher four-class accuracy, positive predictive values above 90, lower computational cost, and fewer training epochs.These goals jointly address performance, model complexity, and training speed.
- Approach: Progressive image resizing, automatic learning-rate finding, and discriminative learning rates are used to improve performance and reduce training time.The authors describe these as state-of-the-art training techniques applied with continuous human input.
A. COVIDx Dataset
COVIDx is a public four-class chest-radiography dataset assembled from two publicly available datasets, containing 5,941 images from 2,839 patients. COVID-19 cases are substantially fewer than the other classes, creating class imbalance.
- Dataset composition: COVIDx contains 5,941 posteroanterior chest radiographs from 2,839 patients across Normal, Bacterial, Viral, and COVID-19 classes.Normal denotes no infection; Viral denotes non-COVID-19 viral pneumonia.
- Dataset composition: The dataset was curated by combining two publicly available datasets and released in a pre-processed form.The paper identifies the COVID-Net repository as the source of the available pre-processed version.
- Class distribution: COVIDx includes 68 COVID-19 radiographs from 45 patients, compared with 1,203 Normal, 931 bacterial-pneumonia, and 660 non-COVID-19 viral-pneumonia patients.The authors characterize this distribution as a class-imbalance problem.
- Class distribution: Figure 2 separates training-image counts from aggregate test-set counts and shows aggregate patient counts for each category.Blue bars represent training data, yellow bars represent the aggregate test set, and Normal means negative infection.
B. Data Augmentation
The study augments training data with random image transformations to create additional examples and support model regularization and generalization.
- Augmentation operations: Training images receive vertical flips, random rotations up to 15 degrees, and lighting-condition transformations.Only the training data was augmented.
C. Network Architecture
COVID-ResNet uses a pretrained 50-layer residual network rather than introducing a new architecture. Its training procedure supports varying image sizes, with a replaced classification head fine-tuned on COVIDx.
- Architecture choice: COVID-ResNet is a 50-layer ResNet50 variant selected for its performance, parameter balance, faster training, and ability to accept different image sizes.The architecture leverages an existing convolutional neural network rather than proposing a new design.
- Input preparation: ResNet50 uses ImageNet-pretrained weights, while COVIDx images are trained at 128x128x3, 224x224x3, and 299x299x3 resolutions.ImageNet channel normalization is applied to match the pretrained network’s input preparation.
- Transfer learning: The transfer-learning head is replaced with adaptive pooling, batch normalization, dropout, and linear layers before COVIDx fine-tuning.This head follows the sequence proposed in the referenced fastai approach.
D. Training the Network
COVID-ResNet is trained through progressive resizing, with learning rates reduced across stages to preserve learned weights and improve generalization. The staged procedure combines head training, whole-network fine-tuning, and progressively larger inputs.
- Progressive resizing: Progressive resizing trains the network on increasingly larger input images across three stages.The approach transfers learning across input image sizes while retaining global features at different pixel resolutions.
- Stage-wise fine-tuning: Stage I first trains only the newly added head, then fine-tunes the entire network using discriminative learning rates.The head is trained for 3 epochs at 1e-3, followed by whole-network fine-tuning for 5 epochs.
- Learning-rate control: Learning rates are reduced in later stages so successive updates modify the weights less substantially.This preserves the learned weights as training moves from one resolution stage to the next.
- Optimization and tooling: Training uses the Adam optimizer with batch size 32, while preprocessing, augmentation, and training are implemented in fastai.
III. RESULTS AND DISCUSSION
COVID-ResNet substantially outperformed the original COVID-Net results on the COVIDx test set while using fewer parameters. The authors caution that perfect COVID-19 metrics are based on a very limited test set, and report improvements across classes, including the Normal class.
- Overall performance: 96.23% accuracy versus 83.5% for COVID-Net, with 25.6M versus 116.6M parameters.The reported comparison represents about a 13% performance improvement and 4.5 times fewer parameters.
- Class-wise evaluation: Tables report sensitivity, positive predictive value, and F1-score for each infection class and the Normal class.The overall confusion matrix is presented in Figure 4.
- Limitations: The model achieved 100% on all reported COVID-19 metrics in this test dataset, but the COVID-19 test set was very small relative to the other classes.The authors plan further testing and improvement as additional COVID-19 data becomes available.
- Class-wise evaluation: Only 24 of 637 test cases were misclassified, with the non-COVID-19 viral class showing lower PPV than the other classes.The authors attribute significant performance improvements to data augmentation relevant to the data.
- Practical implications: Higher sensitivity and PPV on the Normal class may reduce false positives across COVID-19 and the other infection classes.The authors also plan methods to explain model predictions and investigate patterns in radiographs.
IV. CONCLUSIONS
COVID-ResNet classifies COVID-19 and three other infection types on the public COVIDx dataset, while emphasizing efficient training and improved generalization. Despite promising accuracy, it is not intended for direct clinical diagnosis and requires larger-scale validation.
- COVID-ResNet classifies COVID-19 and three other infection types using the publicly available COVIDx dataset.
- Data augmentation increases the training set size and improves generalization.
- Human intervention and state-of-the-art training techniques help identify learning rates that improve model performance and speed.
- COVID-ResNet is not intended for direct clinical diagnosis.Clinical usefulness requires training on a larger dataset and testing in the wild with a larger cohort.