Source-linked AI summary
Automatic Detection of Coronavirus Disease (COVID-19) Using X-ray Images and Deep Convolutional Neural Networks
Ali Narin, Ceren Kaya, Ziynet Pamuk
TL;DR
COVID-19 became a serious worldwide public health problem, while radiologists faced increased workload and chest X-ray offered a fast, cheap, common clinical method. The study compared five CNN models for automatic COVID-19 detection, with ResNet50 achieving the highest accuracy across three datasets.
Problem
COVID-19 became a serious worldwide public health problem, while manual diagnoses increased radiologists' workload.
Method
The study compared five CNN models across three studies and used a method without manual feature extraction, selection, or classification.
Results
ResNet50 yielded the highest accuracy among five models across three datasets: 96.1% for Dataset-1, 99.5% for Dataset-2, and 99.7% for Dataset-3.
Takeaways & Limitations
The findings support automatic COVID-19 detection using chest X-ray images, with ResNet50 performing best among the evaluated models.
Takeaways & Limitations
The study is restricted by limited data; testing data from many different centers is needed to create more stable results.
Abstract
from arXiv · showhide
The 2019 novel coronavirus disease (COVID-19), with a starting point in China, has spread rapidly among people living in other countries, and is approaching approximately 34,986,502 cases worldwide according to the statistics of European Centre for Disease Prevention and Control. There are a limited number of COVID-19 test kits available in hospitals due to the increasing cases daily. Therefore, it is necessary to implement an automatic detection system as a quick alternative diagnosis option to prevent COVID-19 spreading among people. In this study, five pre-trained convolutional neural network based models (ResNet50, ResNet101, ResNet152, InceptionV3 and Inception-ResNetV2) have been proposed for the detection of coronavirus pneumonia infected patient using chest X-ray radiographs. We have implemented three different binary classifications with four classes (COVID-19, normal (healthy), viral pneumonia and bacterial pneumonia) by using 5-fold cross validation. Considering the performance results obtained, it has seen that the pre-trained ResNet50 model provides the highest classification performance (96.1% accuracy for Dataset-1, 99.5% accuracy for Dataset-2 and 99.7% accuracy for Dataset-3) among other four used models.
1. Introduction
COVID-19 created urgent diagnostic pressure, while chest X-ray offered a fast, inexpensive, widely used imaging option whose interpretation requires specialist expertise. The introduction motivates computer-aided and artificial-intelligence methods as potential support for diagnosis and radiologist workload.
- COVID-19 emerged in Wuhan in December 2019 and became a serious worldwide public-health problem.
- Chest X-ray is a fast, cheap, common method used worldwide for pneumonia diagnosis.It also exposes patients to less radiation than CT and MRI.
- Correctly diagnosing pneumonia from X-ray images requires substantial expert knowledge and experience.The paper states that COVID-19 diagnosis from chest X-rays can only be performed by specialist physicians.
- The limited number of specialists and broader shortages of doctors, hospital beds, and health personnel increase diagnostic pressure during the pandemic.The paper also notes that healthcare workers face infection risk because COVID-19 is highly contagious.
- Early pneumonia diagnosis is described as vital for slowing epidemic spread through patient quarantine and supporting patient recovery.
- Computer-aided diagnosis can help doctors diagnose pneumonia more quickly and accurately, while reducing workload and increasing reliability and quantitative analysis.The introduction connects this potential with the ability of artificial intelligence to handle datasets exceeding human capacity.
2. Related Works
Prior studies explored CNNs, feature-based classifiers, segmentation, and transfer learning for COVID-19 detection from X-ray or CT images. This study extends that landscape with end-to-end pre-trained CNN comparisons across multiple classes and datasets.
- Earlier COVID-19 studies used CNNs and transfer learning, while others classified manually extracted image features with SVM, kNN, softmax, or random forest.
- Prior studies varied in dataset size, feature-processing strategy, and cross-validation design, including 2-fold, 5-fold, and 10-fold validation.
- Related work included X-ray and CT classification, infection-area segmentation, and models designed to assist physicians in identifying diseased lung regions.
- Researchers applied diverse architectures including Xception, VGG19, DenseNet, SqueezeNet, ResNet, InceptionV3, InceptionResNet, Bayesian CNNs, and Capsule Networks.
- The authors’ earlier study trained on only 50 normal and 50 COVID-19-positive cases because COVID-19 data were insufficient.
- This study proposes end-to-end pre-trained CNN models without manual feature extraction or selection, comparing five models across COVID-19, normal, bacterial-pneumonia, and viral-pneumonia classes.The paper presents the approach as using more data than many studies and as a potential decision-support system for radiologists.
3.1 Dataset
The dataset combines COVID-19, normal, bacterial-pneumonia, and viral-pneumonia chest X-ray images from multiple repositories. The authors organize these images into three binary datasets for model evaluation.
- The study includes chest X-ray images from 341 COVID-19 patients.
- COVID-19 images were obtained from an open-source GitHub repository containing chest X-ray and CT images from several respiratory conditions.
- The dataset contains 2,800 normal chest X-ray images selected from the ChestX-ray8 database.
- It also uses 2,772 bacterial-pneumonia and 1,493 viral-pneumonia chest X-ray images from the Kaggle Chest X-Ray Images (Pneumonia) repository.
- The images were organized into three binary datasets containing combinations of normal, COVID-19, bacterial-pneumonia, and viral-pneumonia classes.
- All dataset images were resized to 224x224 pixels, with class distributions reported in Table 1.
3.2 Architecture of Deep Transfer Learning
The section describes CNN architectures and their sequential feature-extraction and classification operations, then introduces transfer learning with five pre-trained models for COVID-19 chest X-ray classification.
- Deep learning models are applied to medical-image classification, segmentation, lesion detection, and disease analysis using MRI, CT, and X-ray data.
- CNNs convert input images into matrices, learn image-label differences during training, and use them to predict labels for new images.
- Convolutional and pooling layers extract features, while fully connected layers perform classification.
- 3.2.1 Convolutional Layer: Convolutional layers apply sliding kernels to extract low- and high-level features, with stride controlling movement across the input matrix.
- 3.2.2 Pooling Layer: Pooling reduces feature-map outputs and network parameters; this study uses max-pooling and global average pooling, while dropout targets overfitting and divergence.
- 3.2.3 Fully Connected Layer: The fully connected layer acts like a multilayer perceptron, using ReLU and Softmax activations for processing and output prediction.
- 3.2.4. Pre-Trained Models: The study builds ResNet50, ResNet101, ResNet152, InceptionV3, and Inception-ResNetV2 models for classifying COVID-19 chest X-ray images across three binary comparisons.
- 3.2.4. Pre-Trained Models: Transfer learning uses ImageNet-derived information to train models with fewer datasets and lower calculation costs.
3.3 Experimental Setup
The experiments train five CNN models with fixed optimization settings on resized, randomly split datasets, using 5-fold cross validation to evaluate performance.
- Experiments were performed on Google Colaboratory using Ubuntu 16.04 with CPU, Tesla K80 GPU, or TPU hardware.
- Five CNN models—ResNet50, ResNet101, ResNet152, InceptionV3, and Inception-ResNetV2—were evaluated.
- The models were trained by optimizing cross-entropy with the ADAM optimizer, using β1 = 0.9 and β2 = 0.999.
- All experiments used batch size 3, learning rate 1e-5, and 30 epochs.
- The datasets were randomly split into independent training and testing sets, with 80% used for training and 20% for testing.
- 5-fold cross validation produced results for five k values, k=1-5.
3.4 Performance Metrics
The study evaluates deep transfer-learning models using five performance criteria based on confusion-matrix quantities for COVID-19 and normal classifications.
- Five criteria were used to assess the performance of the deep transfer-learning models.
- Accuracy is calculated as (TN + TP) / (TN + TP + FN + FP).
- Recall is calculated as TP / (TP + FN), while specificity is calculated as TN / (TN + FP).
- Precision is calculated as TP / (TP + FP), and F1-Score is calculated as 2x((PrecisionxRecall)/(Precision+Recall)).
- TP, FP, TN, and FN denote true positive, false positive, true negative, and false negative quantities, respectively.
- For Dataset-1, TP denotes COVID-19 cases correctly labeled as COVID-19, while FP denotes normal cases mislabeled as pneumonia classes.
4. Experimental Results
The study evaluates five pre-trained CNNs across three binary COVID-19 classifications using 5-fold cross-validation. ResNet50 generally performs strongly, while performance varies across datasets and models.
- Experimental setup: Three binary classifications compare COVID-19 with normal, viral pneumonia, and bacterial pneumonia classes using five pre-trained CNN models.The evaluation uses 5-fold cross-validation with 80% of data for training and 20% for testing in each fold.
- Binary Class-1: ResNet50 performs better than the other models during training for the COVID-19/normal classification.Training accuracy and loss are shown in Figures 4 and 5, while testing accuracy is shown in Figure 6.
- Binary Class-1: 96.1% overall performance is reported for ResNet50 and ResNet101 in the COVID-19/normal classification.The authors also state that excess normal data results in higher performance across all models.
- Binary Class-2: ResNet50 and ResNet101 perform better than the other models in training for the COVID-19/viral-pneumonia classification.The corresponding training accuracy and loss are presented in Figures 7 and 8.
- Binary Class-2: 99.4% and 99.5% are reported for COVID-19 detection in the COVID-19/viral-pneumonia classification.The models reach quite high values across each fold, and test-data results are described as generally stable.
- Binary Class-3: 100% COVID-19 detection is reported for InceptionV3 in the COVID-19/bacterial-pneumonia classification, while ResNet50 has high overall performance.ResNet50 also exhibits higher training performance, whereas InceptionV3 improves toward the end of training.
5. Discussion
The discussion positions the study as a larger, end-to-end comparison of five CNNs for distinguishing COVID-19 from other chest conditions. It proposes decision support for radiologists while acknowledging limited data and the need for multi-center testing.
- Method: The method uses raw data end-to-end without manual feature extraction, feature selection, or classification.The paper describes this as a direct end-to-end procedure.
- Findings: The reported COVID-19 classification performance is significantly higher across normal, viral-pneumonia, and bacterial-pneumonia classes.This is presented as a principal study finding alongside the comparison of five CNN models.
- Clinical relevance: The authors propose a high-accuracy decision-support system for radiologists performing automatic COVID-19 detection and patient follow-up.The motivation includes increased radiologist workload and possible diagnostic errors associated with fatigue.
- Limitations: Limited data constrain the study, and testing with data from many centers is identified as necessary for more stable results.The paper also lists multi-center evaluation as a future direction.
6. Conclusion
The study uses deep transfer learning with chest X-ray images to automatically detect COVID-19 across several pneumonia and normal classes. Among five evaluated CNN models and three binary datasets, ResNet50 achieved the highest reported accuracy, while the authors identify further performance evaluation as a future direction.
- Early COVID-19 prediction is presented as important for preventing disease spread.
- The proposed deep transfer learning approach uses chest X-ray images from normal, COVID-19, bacterial pneumonia, and viral pneumonia patients.
- Five CNN models were studied and compared for COVID-19 detection.
- 96.1% accuracy for Dataset-1, 99.5% accuracy for Dataset-2, and 99.7% accuracy for Dataset-3 were achieved by the ResNet50 pre-trained model as the highest accuracy among the five models.
- The findings are described as potentially helping radiologists make clinical decisions and providing insight into early COVID-19 detection with deep transfer learning.
- The authors identify improving classification performance through different approaches in subsequent studies as a future direction.