Source-linked AI summary
Classification of COVID-19 in chest X-ray images using DeTraC deep convolutional neural network
Asmaa Abbas, Mohammed M. Abdelsamea, Mohamed Medhat Gaber
TL;DR
COVID-19 chest X-ray classification faces limited annotated medical images and irregular data distributions. The paper adapts DeTraC, combining transfer learning with class decomposition, and reports 95.12% accuracy on a comprehensive chest X-ray dataset.
Problem
Limited annotated medical images and irregularities in medical-image datasets challenge COVID-19 chest X-ray classification.
Method
DeTraC adapts a pre-trained CNN by adding class decomposition, treating subclasses independently before composing predictions for the original classes.
Results
95.12% accuracy was achieved by DeTraC with ResNet on chest X-ray images.
Takeaways & Limitations
DeTraC demonstrated robustness to limited training images and irregular data distributions in COVID-19 chest X-ray classification.
Takeaways & Limitations
The authors identify larger-dataset validation, explainability, and deployment efficiency as future work.
Abstract
from arXiv · showhide
Chest X-ray is the first imaging technique that plays an important role in the diagnosis of COVID-19 disease. Due to the high availability of large-scale annotated image datasets, great success has been achieved using convolutional neural networks (CNNs) for image recognition and classification. However, due to the limited availability of annotated medical images, the classification of medical images remains the biggest challenge in medical diagnosis. Thanks to transfer learning, an effective mechanism that can provide a promising solution by transferring knowledge from generic object recognition tasks to domain-specific tasks. In this paper, we validate and adapt our previously developed CNN, called Decompose, Transfer, and Compose (DeTraC), for the classification of COVID-19 chest X-ray images. DeTraC can deal with any irregularities in the image dataset by investigating its class boundaries using a class decomposition mechanism. The experimental results showed the capability of DeTraC in the detection of COVID-19 cases from a comprehensive image dataset collected from several hospitals around the world. High accuracy of 95.12% (with a sensitivity of 97.91%, a specificity of 91.87%, and a precision of 93.36%) was achieved by DeTraC in the detection of COVID-19 X-ray images from normal, and severe acute respiratory syndrome cases.
1 Introduction
Chest X-ray supports COVID-19 diagnosis, while CNNs and transfer learning offer approaches for medical-image classification with limited annotation. DeTraC adapts class decomposition to address irregularities in chest X-ray data.
- Chest X-ray is an important imaging technique for diagnosing COVID-19 and can distinguish normal, COVID-19, and SARS cases.
- CNNs automatically learn features from domain-specific images, unlike classical machine-learning methods.
- Transfer learning transfers knowledge from a pre-trained network to a new task, reducing the need for large annotated datasets.
- Class decomposition partitions classes into simpler subclasses to cope with irregular data distributions and improve classification flexibility.
- DeTraC adds a class decomposition layer to pre-trained models, treating subclasses independently before composing final predictions.
2 Related work
Prior work applied machine learning, CNNs, transfer learning, and imaging data to COVID-19 detection. The paper identifies data irregularities as an insufficiently explored challenge affecting classification accuracy.
- Fast detection of COVID-19 through genetic and imaging tests may contribute to controlling disease spread.
- CNNs and related approaches have been applied to COVID-19 detection in chest X-ray and CT images.
- Prior chest X-ray systems combined ImageNet-pre-trained CNN features with SVM or used transfer-learning architectures for multiple disease classes.
- Despite deep-learning success, prior COVID-19 detection work had not explored data irregularities such as overlapping classes.
- Data irregularities can affect the resulting accuracy of machine-learning models, motivating the focus of this work.
3 DeTraC method
DeTraC combines pretrained CNN feature extraction with class decomposition and composition to classify COVID-19 chest X-ray images. Its workflow transforms original classes into homogeneous subclasses for training, then recomposes them for final predictions.
- 3.1 DeTraC architecture overview: DeTraC uses three phases: pretrained CNN feature extraction, class decomposition with gradient-based training, and class composition for final classification.Decomposition and composition are placed before and after knowledge transfer from an ImageNet-pretrained CNN.
- 3.2 Feature-space construction: PCA projects high-dimensional CNN features into a lower-dimensional space before decomposition, ignoring highly correlated features to form more homogeneous classes.The paper states that this also reduces memory requirements and improves framework efficiency.
- 3.3 Class decomposition: The feature space is represented as dataset A with class labels L, then mapped to dataset B while preserving instances and replacing labels with subclass labels C.The relationship is expressed as A = (A|L) 7→ B = (B|C).
- 3.3 Class decomposition: K-means assigns patterns to subclasses using the nearest centroid under squared Euclidean distance, producing dataset B with new subclass labels.Each original class is further divided into k subclasses after clustering.
- 3.2 Class decomposition and composition: Class decomposition partitions each original image class into several independent subclasses, while composition assembles those subclasses into predictions for the original classes.The decomposition layer is intended to simplify the local structure of the data distribution.
- 3.5 Evaluation and composition: Evaluation uses multiclass confusion matrices and Accuracy, Specificity, and Sensitivity, with TP, TN, FP, and FN defined for COVID-19 and other cases.The method is experimentally validated for COVID-19 detection from chest X-ray images.
4 Experimental study
The experimental study evaluates DeTraC on combined chest X-ray datasets using augmentation, transfer learning, and class decomposition. Performance is assessed through classification metrics, learning curves, and ROC analysis.
- Dataset and preprocessing: The study combines normal, COVID-19, and SARS chest X-ray samples from two datasets and applies augmentation and contrast enhancement.The augmented dataset contains 1764 samples.
- DeTraC configuration: AlexNet extracts features, while k-means clustering with k = 2 decomposes each original class into subclasses.The decomposed classes are used within the DeTraC pipeline.
- Parameter settings: The dataset is split into 70% training and 30% evaluation, with ResNet18 used as the ImageNet-pre-trained transfer-learning network.The final fully connected layer is changed for the new classification task.
- Performance analysis: Training and test accuracy and error are plotted as learning curves, and the area under the receiver operating characteristic curve is computed.The ROC analysis is presented for the DeTraC model trained with a ResNet pre-trained network.
- Accuracy evaluation: DeTraC-ResNet18 is compared with ResNet18 under the same settings to evaluate robustness in COVID-19 image classification.The reported ResNet18 accuracy is 92.5%, with sensitivity of 65.01%; the specificity value is not complete in the supplied passage.
5 Discussion
The discussion frames limited annotated medical data and irregular class distributions as challenges for CNN-based COVID-19 classification. It presents transfer learning and DeTraC’s class decomposition as responses to these challenges.
- Data and training challenges: Limited availability of annotated medical images makes end-to-end CNN training impractical for medical imaging.Transfer learning addresses this setting by transferring knowledge from pre-trained CNNs to medical tasks.
- Data irregularities: Data irregularities can miscalibrate class boundaries and affect classification accuracy, especially when classes overlap or differ in representation.The discussion identifies irregularities as an insufficiently explored issue in COVID-19 chest X-ray detection.
- Reported outcome: 95.12% accuracy was achieved by DeTraC with ResNet on chest X-ray images, while the method is described as robust to limited training images and irregular data distributions.The discussion also describes the class decomposition layer as a generic solution intended to improve CNN efficiency.
6 Conclusion and future work
The conclusion addresses irregularities in COVID-19 chest X-ray data by adapting DeTraC with class decomposition. Future work will test larger datasets and improve explainability and deployment efficiency.
- Conclusion: Irregularities in annotated chest X-ray data are identified as the central challenge for COVID-19 case classification.The paper places this challenge within the broader use of CNNs for medical image classification.
- Conclusion: DeTraC is adapted as a class-decomposition deep CNN for COVID-19 image classification in a comprehensive chest X-ray dataset.The conclusion describes the approach as effective and robust for data irregularity and limited training images.
- Future work: Future experiments will validate the method on larger datasets and add explainability, pruning, and quantisation for usability and handheld deployment.These directions are presented as planned extensions rather than completed evaluations.