Source-linked AI summary
Automatic Detection of Coronavirus Disease (COVID-19) in X-ray and CT Images: A Machine Learning-Based Approach
Sara Hosseinzadeh Kassani, Peyman Hosseinzadeh Kassasni, Michal J. Wesolowski, Kevin A. Schneider, Ralph Deters
TL;DR
The paper addresses automatic COVID-19 classification from chest X-ray and CT images, motivated by the disease’s transmissibility and the need for early detection and monitoring. It compares pre-trained CNN feature extractors combined with machine-learning classifiers, avoiding task-specific preprocessing. DenseNet121 with Bagging achieved 99% classification accuracy, while ResNet50 with LightGBM achieved 98%; the study notes limited training data and difficulty with challenging images.
Problem
The study addresses the need for computer-aided COVID-19 detection and monitoring from chest X-ray and CT images.
Method
Pre-trained CNNs extract image feature vectors that are supplied to machine-learning classifiers for COVID-19 classification.
Results
99% classification accuracy was achieved by the DenseNet121 feature extractor with a Bagging tree classifier, while ResNet50 with LightGBM achieved 98%.
Takeaways & Limitations
The feature-extractor and classifier approach produced strong COVID-19 classification performance with low extraction and training time compared with training a deep CNN from scratch.
Takeaways & Limitations
The study is limited by training-data size and by challenging cases with vague, low-contrast boundaries and artifacts.
Abstract
from arXiv · showhide
The newly identified Coronavirus pneumonia, subsequently termed COVID-19, is highly transmittable and pathogenic with no clinically approved antiviral drug or vaccine available for treatment. The most common symptoms of COVID-19 are dry cough, sore throat, and fever. Symptoms can progress to a severe form of pneumonia with critical complications, including septic shock, pulmonary edema, acute respiratory distress syndrome and multi-organ failure. While medical imaging is not currently recommended in Canada for primary diagnosis of COVID-19, computer-aided diagnosis systems could assist in the early detection of COVID-19 abnormalities and help to monitor the progression of the disease, potentially reduce mortality rates. In this study, we compare popular deep learning-based feature extraction frameworks for automatic COVID-19 classification. To obtain the most accurate feature, which is an essential component of learning, MobileNet, DenseNet, Xception, ResNet, InceptionV3, InceptionResNetV2, VGGNet, NASNet were chosen amongst a pool of deep convolutional neural networks. The extracted features were then fed into several machine learning classifiers to classify subjects as either a case of COVID-19 or a control. This approach avoided task-specific data pre-processing methods to support a better generalization ability for unseen data. The performance of the proposed method was validated on a publicly available COVID-19 dataset of chest X-ray and CT images. The DenseNet121 feature extractor with Bagging tree classifier achieved the best performance with 99% classification accuracy. The second-best learner was a hybrid of the a ResNet50 feature extractor trained by LightGBM with an accuracy of 98%.
1. Introduction
COVID-19 is highly transmissible and can cause severe pneumonia, creating a need for early detection and progression monitoring. The study motivates computer-aided analysis of chest X-ray and CT images and surveys deep-learning feature extraction with machine-learning classifiers.
- COVID-19 emerged from pneumonia cases identified in Wuhan in December 2019 and is caused by SARS-CoV-2.
- Medical imaging can support confirmation of COVID-19 pneumonia and monitoring, although it is not recommended for primary diagnosis in Canada.Chest X-ray and CT images show ground-glass opacity abnormalities whose prominence varies across illness days.
- Computer-aided diagnosis systems using image-processing techniques and deep learning could assist diagnosis and understanding of disease progression.
- The study aims to provide a generic feature-extraction method without handcrafted or complex features for application to X-ray and CT images.The stated goals also include reducing generalization error and improving diagnostic accuracy.
- The approach uses transfer learning without data augmentation or extensive preprocessing to address limited training images and improve generalization.
- A deep CNN combined with Bagging trees achieved very good COVID-19 classification performance despite limited image samples, with a stated aim of reducing detection time.The paper also reports an end-to-end web-based system for screening suspicious cases.
2. Proposed Methodology
The methodology uses pre-trained CNNs to encode X-ray and CT images as low-dimensional feature vectors, then supplies those representations to machine-learning classifiers. Transfer learning addresses insufficient labeled data and the risk of overfitting.
- The pipeline collects public chest X-ray and CT images and applies standard image normalization before feature extraction.
- Transfer learning reuses weights learned in an auxiliary domain to address insufficiently labeled data and overly deep CNN models.
- CNN Descriptor: Pre-trained CNNs encode input images into sparse, low-dimensional feature vectors that are then fed to machine-learning classifiers.
3. Experiments
The experiments use a balanced public dataset of COVID-19 and healthy X-ray and CT images, with limited preprocessing before feature extraction and standard classification metrics for evaluation.
- Dataset description: 137 COVID-19-positive images were combined with 137 healthy images to balance X-ray and CT classes.The positive set contained 117 chest X-ray and 20 CT images, matched with 117 healthy X-ray and 20 healthy CT images.
- Method framework: The framework includes a VGGNet feature-extraction configuration for processing the medical images.
- Dataset description: Confirmed COVID-19 X-rays show varied opacity patterns, including bilateral lower-lung infiltrates, infrahilar opacities, streaky basal opacities, and lobar opacities.One case also shows increasing opacity density over time, while another is described on day 7 after symptom onset.
- Data pre-processing: The method avoids extensive preprocessing to improve robustness to noise, artifacts, and intensity variations across imaging clinics and equipment.
- Data pre-processing: Images were resized to 600×450 pixels, with architecture-specific sizes of 331×331 for NASNetLarge and 224×224 for NASNetMobile.
- Evaluation criteria: Evaluation uses recall, precision, accuracy, and F1-score computed from true-positive, false-negative, true-negative, and false-positive counts.Recall measures correctly classified COVID-19 cases; precision, accuracy, and F1-score provide complementary classification measures.
4. Discussion
The study evaluates deep CNN feature extractors paired with machine-learning classifiers for COVID-19 classification from chest X-ray and CT images. DenseNet121 with Bagging achieved the strongest accuracy, while results also reveal a substantial speed–accuracy trade-off and limitations on difficult cases and clinical validation.
- Classification performance: 99.00% ± 0.09 accuracy was achieved by DenseNet121 features with the Bagging classifier, the best result among the evaluated combinations.The comparison used six machine-learning algorithms applied to features extracted by deep CNNs.
- Classification performance: 98.00 ± 0.09 accuracy was obtained by ResNet50 with LightGBM, making it the second-best combination.DenseNet121 with Bagging was reported as 1% better in classification accuracy.
- Efficiency and stability: ResNet50 with LightGBM required almost 30 times less training time than DenseNet121 with Bagging, despite 1% lower accuracy.The authors describe Bagging as slower but more stable because it had the lowest standard deviation.
- Classification performance: 76.00 ± 0.32 accuracy was the worst reported result, obtained with ResNet101V2 features trained by AdaBoost.Figures 5 and 6 summarize classifier and feature-extractor accuracy distributions from 10-fold cross-validation.
- Additional metrics: Bagging produced 99.00% precision, recall, and F-score for MobileNet and InceptionV3 feature vectors.XGBoost and Bagging also produced 98.00% precision, recall, and F-score with DenseNet121, DenseNet201, and Xception features.
- Limitations and applicability: The approach requires minimal preprocessing and no handcrafted features, but limited training data, input rescaling, difficult images, artifacts, and absent real clinical trials constrain interpretation.The authors specifically call for evaluation of diagnostic performance in real clinical study trials.
5. Conclusion
The paper presents a feature extractor-based deep learning and machine learning classifier approach for computer-aided diagnosis of COVID-19 pneumonia, evaluated on available chest X-ray and CT data.
- The study trains several machine learning algorithms on features extracted by established CNN architectures to identify an effective feature–learner combination.