Source-linked AI summary
Automated diagnosis of COVID-19 with limited posteroanterior chest X-ray images using fine-tuned deep neural networks
Narinder Singh Punn, Sonali Agarwal
TL;DR
Limited COVID-19-positive imaging samples complicate unbiased automated diagnosis from chest X-rays. The paper uses oversampling, weighted loss, and transfer learning across several deep networks for binary and multiclass classification; NASNetLarge generally performs best, especially for binary classification, while visual explanations show prediction bases.
Problem
Public COVID-19 imaging datasets contain limited positive samples, creating a challenge for unbiased learning.
Method
The study applies random oversampling and weighted class loss with fine-tuned transfer learning in several deep-learning architectures for chest X-ray classification.
Results
NASNetLarge generally achieves better scores than the other architectures, particularly for binary COVID-19 classification, although performance depends on the scenario.
Takeaways & Limitations
Binary classification groups other diseases as non-COVID-19, and visual explanations are used to show the basis of model predictions.
Takeaways & Limitations
The article reports no studies involving human participants or animals.
Abstract
from arXiv · showhide
The novel coronavirus 2019 (COVID-19) is a respiratory syndrome that resembles pneumonia. The current diagnostic procedure of COVID-19 follows reverse-transcriptase polymerase chain reaction (RT-PCR) based approach which however is less sensitive to identify the virus at the initial stage. Hence, a more robust and alternate diagnosis technique is desirable. Recently, with the release of publicly available datasets of corona positive patients comprising of computed tomography (CT) and chest X-ray (CXR) imaging; scientists, researchers and healthcare experts are contributing for faster and automated diagnosis of COVID-19 by identifying pulmonary infections using deep learning approaches to achieve better cure and treatment. These datasets have limited samples concerned with the positive COVID-19 cases, which raise the challenge for unbiased learning. Following from this context, this article presents the random oversampling and weighted class loss function approach for unbiased fine-tuned learning (transfer learning) in various state-of-the-art deep learning approaches such as baseline ResNet, Inception-v3, Inception ResNet-v2, DenseNet169, and NASNetLarge to perform binary classification (as normal and COVID-19 cases) and also multi-class classification (as COVID-19, pneumonia, and normal case) of posteroanterior CXR images. Accuracy, precision, recall, loss, and area under the curve (AUC) are utilized to evaluate the performance of the models. Considering the experimental results, the performance of each model is scenario dependent; however, NASNetLarge displayed better scores in contrast to other architectures, which is further compared with other recently proposed approaches. This article also added the visual explanation to illustrate the basis of model classification and perception of COVID-19 in CXR images.
1 Introduction
COVID-19 diagnosis requires accessible methods to identify lung infection and distinguish COVID-19 from other pneumonia, especially given limited positive samples. The section motivates deep-learning analysis of chest X-rays and introduces imbalance-aware transfer learning.
- COVID-19 is highly contagious, making identification of infected people important for controlling spread.
- Chest X-rays are preferred over CT because they are more widely available, faster, and cheaper.CXR can be used to analyze lung abnormalities associated with pneumonia and inflammation.
- Deep-learning methods, particularly convolutional neural networks, are effective for feature extraction and learning in medical image analysis.
- The study uses COVID-19 and RSNA chest X-ray images to generate samples and support comparison with other research.
- Limited COVID-19-positive samples may bias learning, motivating random oversampling and weighted class loss for fine-tuned transfer learning.
2 Related work
Prior work applies deep learning, transfer learning, augmentation, GANs, and specialized architectures to automated pneumonia and COVID-19 analysis from chest X-rays. This paper extends that direction with state-of-the-art architectures and class-imbalance learning.
- Researchers have used chest X-ray deep-learning systems for pneumonia detection and COVID-19 prediction.
- Transfer learning with pretrained models such as ResNet50, InceptionV3, and Inception-ResNetV2 has been applied to COVID-19 chest X-ray prediction.
- GAN-based augmentation has been used to generate additional chest X-ray images while addressing overfitting in pneumonia classification.
- Other studies proposed COVIDResNet with progressive resizing and adaptive learning-rate techniques for multiclass classification.
- This article combines state-of-the-art deep-learning architectures with transfer learning and class-imbalance learning for early COVID-19 diagnosis.
3 Dataset description
The study combines three publicly available chest X-ray datasets covering COVID-19-related diseases, pneumonia findings, and tuberculosis or normal cases. These sources are assembled to support classification experiments and comparative evaluation.
- The experiments use COVID-19 image, RSNA, and NLM(MC) datasets.
- The COVID-19 image dataset contains chest X-rays associated with COVID-19, MERS, SARS, Streptococcus, Pneumocystis, and ARDS.
- The RSNA dataset contains pneumonia-positive examinations alongside no-finding and other-symptom cases, with expert annotation in two stages.
- RSNA training and testing partitions changed between stages, ultimately producing 26,684 training images and 3,000 test radiographs.
- The NLM(MC) dataset contributes 138 posterior-anterior chest X-rays covering tuberculosis and normal cases.
4 Proposed contribution
The study combines fused CXR datasets, preprocessing, imbalance-handling methods, and fine-tuned deep learning architectures for COVID-19 classification. It supports binary and multi-class strategies while addressing limited COVID-19 samples and image artifacts.
- Proposed contribution: The proposed pipeline uses COVID-19, RSNA, and NLM(MC) datasets to construct binary and multi-class CXR classification problems.Binary classification separates COVID-19 from non-COVID-19 cases; multi-class classification distinguishes COVID-19, pneumonia or tuberculosis, and normal cases.
- Class balancing: Weighted class loss assigns larger weights to minority classes, with class weights computed from the number of classes and samples per class.The weights are fused into the loss function to penalize false predictions involving minority coronavirus samples more heavily.
- Class balancing: Random oversampling increases minority-class samples until class counts are equal, reducing imbalance in the training data.The procedure fills the difference between majority and minority counts using randomly selected minority samples.
- Data preprocessing: The preprocessing pipeline removes textual and symbolic artifacts through threshold-based masking and inpainting, resizes images to 331x331x3, and applies adaptive total variation denoising.The denoising step is described as preserving the original pixel-value distribution while reducing uncertainty and irregular intensities.
- Fine tuning: Transfer learning fine-tunes high-level layers of pretrained architectures while retaining lower-level features learned from earlier convolutional layers.The models discussed include Inception-v3 and Inception-ResNet-v2, whose pretrained knowledge is used to reduce extensive training and computational demands.
5 Evaluation metrics
The models are evaluated on the test set using standard classification metrics, with confusion-matrix-based performance evaluation shown in Fig. 5.
- The trained models are evaluated on test data using accuracy, precision, recall, AUC, specificity, and F1 score.
- Fig. 5 presents the confusion matrix and performance evaluation metrics.
6 Training and testing
Training and validation use several regularization and optimization choices, while performance is monitored across epochs before test-set evaluation.
- The approach uses Adam mini-batch gradient descent on an NVIDIA Titan GPU with batch size 10.
- Training includes 4-fold cross-validation and switch normalization to adapt instance, batch, or layer normalization.
- Early stopping halts training when validation error stops decreasing.
7 Results and discussion
Performance varies across models and scenarios, but NASNetLarge most often achieves the best scores, while binary classification outperforms multi-class classification.
- NASNetLarge mostly achieves the best accuracy, precision, recall, specificity, AUC, and F1-score across the evaluated scenarios.
- Binary COVID-19 versus non-COVID-19 classification produces better results than multi-class classification involving other classes.
8 Visualization
The study combines performance reporting with training curves and visual explanation methods to show how NASNetLarge classifies chest X-ray images.
- Visualization: Visualization is motivated by the need to understand what triggers the model's output rather than treating it as a black box.
- Visualization: NASNetLarge training is monitored with epoch-wise curves for accuracy, AUC, specificity, F1-score, precision, and recall.
- Comparison: The proposed approach is compared with recently proposed classification strategies in Table 5.
- Visualization: LIME explanations display prediction probabilities for normal, COVID-19, and other-pneumonia samples.
- Visualization: CAM and LIME are used to illustrate the image features and patterns influencing COVID-19 classification.
- Visualization: CAM examples cover normal, COVID-19, and other-pneumonia classifications.
- Visualization: Fine-tuned NASNetLarge generates LIME explanations for test samples, while CAM presents learned classification patterns.
9 Conclusion
The study evaluates deep learning models for COVID-19 diagnosis from limited posteroanterior chest X-ray images. Model performance varies by scenario, with NASNetLarge performing especially well in binary COVID-19 classification, while visual explanations clarify prediction bases.
- Models were evaluated across four imbalanced-learning and classification scenarios using accuracy, precision, recall, AUC, specificity, and F1 score.
- NASNetLarge displayed better performance than the other evaluated models, especially for binary classification of COVID-19 samples.
- Local interpretable model-agnostic explanations were used to show the basis of model predictions.
- The authors suggest exploring additional deep learning models and preprocessing techniques to improve results.
Compliance with Ethical Standards
The paper reports no conflicts of interest and no studies involving human participants or animals. It also acknowledges institutional support and identifies an available arXiv preprint.
- The authors declare no conflict of interest.
- The study contains no research involving human participants or animals performed by the authors.
- The authors acknowledge support from IIITA, the Big Data Analytics lab, and colleagues, and note that the manuscript is available as an arXiv preprint.