Source-linked AI summary
A Review on Deep Learning Techniques for the Diagnosis of Novel Coronavirus (COVID-19)
Md. Milon Islam, Fakhri Karray, Reda Alhajj, Jia Zeng
TL;DR
COVID-19 diagnosis faced constrained RT-PCR screening, including resource requirements and high false-negative rates. This paper reviews deep-learning diagnosis systems using CT and X-ray imaging, organizing their datasets, partitioning methods, metrics, challenges, and future directions; 45 systems were reviewed across pre-trained and customized architectures.
Problem
RT-PCR screening is limited by resource demands, strict testing requirements, and high false-negative rates, motivating examination of alternative diagnostic systems.
Method
The paper reviews deep-learning COVID-19 diagnosis systems using CT and X-ray images, presenting a taxonomy and examining datasets, data splitting, evaluation metrics, challenges, and future work.
Results
45 systems were reviewed: 23 used pre-trained models and 22 used customized frameworks; 25 used X-ray data and 20 used CT data.
Takeaways & Limitations
The review provides an organized account of deep-learning COVID-19 diagnosis research across imaging modalities, architectures, datasets, and evaluation practices.
Takeaways & Limitations
COVID-19 imaging research is constrained by a lack of standard data, incomplete or noisy images, ambiguous cases, inaccurate labels, and dataset imbalance.
Abstract
from arXiv · showhide
Novel coronavirus (COVID-19) outbreak, has raised a calamitous situation all over the world and has become one of the most acute and severe ailments in the past hundred years. The prevalence rate of COVID-19 is rapidly rising every day throughout the globe. Although no vaccines for this pandemic have been discovered yet, deep learning techniques proved themselves to be a powerful tool in the arsenal used by clinicians for the automatic diagnosis of COVID-19. This paper aims to overview the recently developed systems based on deep learning techniques using different medical imaging modalities like Computer Tomography (CT) and X-ray. This review specifically discusses the systems developed for COVID-19 diagnosis using deep learning techniques and provides insights on well-known data sets used to train these networks. It also highlights the data partitioning techniques and various performance measures developed by researchers in this field. A taxonomy is drawn to categorize the recent works for proper insight. Finally, we conclude by addressing the challenges associated with the use of deep learning methods for COVID-19 detection and probable future trends in this research area. This paper is intended to provide experts (medical or otherwise) and technicians with new insights into the ways deep learning techniques are used in this regard and how they potentially further works in combatting the outbreak of COVID-19.
I. INTRODUCTION
COVID-19 spread rapidly worldwide, while asymptomatic cases, limited RT-PCR resources, strict testing requirements, and false negatives hinder timely screening. The review examines deep learning systems using CT and X-ray imaging and organizes their datasets, partitioning methods, metrics, challenges, and future directions.
- COVID-19 caused rapidly increasing infections and deaths, widespread quarantine, and substantial disruption to industrial sectors.
- Asymptomatic cases and RT-PCR limitations, including resource constraints, strict testing environments, and high false-negative rates, restrict fast screening.
- CT and X-ray are prominent imaging modalities for COVID-19 diagnosis, with CT preferred over X-ray when available because of its versatility.
- Deep learning support systems have been developed for COVID-19 diagnosis using both CT and X-ray samples.
- The review classifies recent systems, examines experimental data, splitting techniques, and evaluation metrics, and discusses challenges and future research.
II. TAXONOMY OF DEEP LEARNING BASED COVID-19 DIAGNOSIS SYSTEMS
Deep learning COVID-19 diagnosis systems follow a pipeline from imaging-data collection and preparation through partitioning, model training, feature extraction, classification, and evaluation. The taxonomy organizes reviewed systems by learning technique and imaging modality.
- COVID-19 diagnosis pipelines collect CT or X-ray samples, preprocess them, partition data, train models, classify cases, and evaluate performance.
- Data preparation may include noise removal, resizing, and augmentation, while partitioning commonly separates training, validation, and testing data.
- Deep learning automatically extracts features through repeated operations before assigning class labels such as healthy or COVID-19.
- The review covers 45 systems: 23 used pre-trained models and 22 used custom deep learning techniques.
- X-ray images were used by 25 reviewed systems, whereas 20 systems used CT scans.
III. PRE-TRAINED MODEL WITH DEEP TRANSFER LEARNING
Pre-trained models and deep transfer learning reuse knowledge from models trained on related fields, reducing the computational burden and time required to train large models from scratch.
- A pre-trained model has already been trained in a field similar to the application context.
- Transfer learning transfers weights and biases from a large trained model to a similar model for testing or retraining.
- Training a model from scratch on a large dataset requires high computing power and is time consuming.
A. Diagnosis Using Computer Tomography (CT) Images
The reviewed CT-based systems span multi-source and single-source studies using pre-trained or customized architectures, varied datasets, and multiple partitioning strategies. Reported results range across CNN, U-Net, and related designs, with performance measured using accuracy, sensitivity, specificity, AUC, precision, and F1-score.
- Diagnosis Based on Multiple Source Data: Multi-source CT studies used hospital and public data, including ResNet50, ResNet152, 3D U-Net++, and COVNet architectures.
- Diagnosis Based on Multiple Source Data: Xu et al.’s Resnet18 system classified COVID-19, Influenza-A viral pneumonia, and normal cases with 86.7% accuracy, 81.5% sensitivity, 80.8% precision, and 81.1% F1-score.
- Diagnosis Based on Multiple Source Data: COVNet used 4536 CT samples and achieved 90% sensitivity, 96% specificity, and 96% AUC for COVID-19 cases.
- Diagnosis Based on Multiple Source Data: CovidCTNet used 89,145 CT images and achieved 91.66% accuracy, 87.5% sensitivity, 94% specificity, and 95% AUC.
- Diagnosis Based on Single Source Data: Single-source CT studies evaluated multiple CNN variants, UNet++, and transfer-learning models across datasets ranging from 1020 to 46,096 images.
- Diagnosis Based on Single Source Data: UNet++-based detection achieved 94.34% sensitivity, 99.16% specificity, 98.85% accuracy, 88.37% precision, and 99.61% NPV.
B. Diagnosis Using X-ray Images
The review surveys X-ray COVID-19 diagnosis systems using pre-trained CNNs and transfer learning, comparing datasets, partitions, architectures, and evaluation results. Across reviewed scenarios, performance varies by class configuration and model.
- Pre-trained models: Transfer-learning systems use pre-trained CNNs including VGG, ResNet, Inception, Xception, MobileNet, DenseNet, and related architectures for X-ray diagnosis.The reviewed systems draw on multiple open-access and clinical datasets with binary and multiclass labels.
- Datasets and classes: The reviewed X-ray studies use binary and multiclass settings, including COVID-19 versus non-COVID-19, pneumonia, healthy, and SARS categories.Dataset sizes and sources range from small collections to thousands of images assembled from open-access repositories and clinical settings.
- Evaluation: Evaluation measures across the reviewed systems include accuracy, sensitivity, specificity, precision, F1-score, AUC, negative predictive value, and kappa.The review presents these metrics as key factors for comparing diagnostic systems.
- Reported performance: 100% accuracy was reported by GoogLeNet in the two-class scenario, compared with 85.2% for AlexNet in the three-class scenario and 80.6% for GoogLeNet in the four-class scenario.These results come from experiments conducted under different class-level dataset scenarios.
- Reported performance: ResNet50 achieved Accuracy=98.18, Sensitivity=98.24, Precision=98.14, and F1-Score=98.19 on a three-class X-ray diagnosis task.The experiment used 278 images containing COVID-19, healthy, and pneumonia cases.
IV. CUSTOM DEEP LEARNING TECHNIQUES
Custom deep learning techniques tailor architectures to the COVID-19 diagnosis task through specialized methods, hybrid deep-learning designs, or combinations with other AI fields.
- Custom architectures: Custom networks allow architectures to be designed for a specific application of interest.The paper associates this customization with the potential for more consistent and accurate performance, while describing it as an architectural capability.
- Hybrid designs: Custom techniques may use a particular deep learning method, hybridize deep learning algorithms, or combine deep learning with machine learning, data mining, and nature-inspired algorithms.These categories define the main forms of customization discussed in the section.
A. Diagnosis Using Computer Tomography (CT) Images
The review covers customized deep-learning systems for CT-based COVID-19 diagnosis, including CNN-based classification, optimization-assisted methods, hybrid feature extraction, and 3D architectures. Reported studies use varied datasets and partitioning strategies.
- Multiple-source data: Modified-Inception reduces attribute dimensionality before final classification in a CNN-based CT diagnosis system.The study randomly partitioned 1040 CT images into training, testing, and validation sets.
- Review organization: The reviewed CT systems use varied architectures and data sources, including customized CNNs, decoder-based designs, and datasets from hospitals and benchmark repositories.Table III summarizes sources, sample counts, classes, partitioning techniques, diagnostic techniques, and performance measures for customized CT systems.
- Single-source data: Farid et al. classified COVID-19 and SARS CT images using hybrid composite feature extraction and 10-fold cross-validation.The dataset contained 51 images in each class, and the reported accuracy, precision, F1-score, and AUC were 94.11%, 99.4%, 94%, and 99.4%, respectively.
B. Diagnosis Using X-ray Images
Customized X-ray systems combine deep neural networks with specialized architectures or machine-learning components for COVID-19 diagnosis. The reviewed studies report varied datasets, class definitions, and performance measures.
- Review organization: Customized X-ray systems include capsule networks, hybrid CNN-machine-learning models, and other specialized architectures summarized by dataset, classes, partitioning, and metrics.Table IV organizes the reviewed customized X-ray diagnosis systems using these factors.
- Datasets and classes: The reviewed X-ray experiments use datasets such as COVID-19 X-ray repositories, RSNA Pneumonia Detection Challenge data, and COVIDx.Class configurations include COVID-19 versus non-COVID-19 and three-class settings with non-COVID pneumonia or healthy cases.
- Single-source data: A hybrid X-ray method combines CNN-based feature extraction and classification with SVM or random-forest classification components.The study used 71 images divided into COVID-19-positive and COVID-19-negative cases.
V. OPEN DISCUSSIONS, CHALLENGES AND FUTURE TRENDS
This section discusses the reviewed COVID-19 diagnosis systems, their challenges, and possible future trends in deep learning-based detection.
- Open Discussions: The section discusses reviewed deep learning systems for COVID-19 diagnosis.
- Challenges: It addresses challenges associated with applying deep learning to COVID-19 detection.
- Future Trends: It considers possible future trends for deep learning-based COVID-19 diagnosis systems.
A. Open Discussions
The review covers 45 systems using CT and X-ray data, comparing architectures, datasets, partitioning strategies, and reported performance while noting important comparability limits.
- Review Scope: 45 systems were reviewed: 23 used pre-trained models and 22 used custom deep learning frameworks.
- Imaging Modalities: 25 systems used X-ray data and 20 used CT data, with most systems relying on multiple data sources.
- CT-Based Systems: CT studies mainly used hospital data from China, while only a few used benchmark datasets or single-source data.
- CT-Based Systems: CT studies mostly used binary classification, and hold-out splitting was more common than cross-validation or random partitioning.
- CT-Based Systems: 99.51% highest accuracy was reported for a pre-trained model, while 99.68% was reported for a customized network in CT diagnosis.
- X-ray-Based Systems: X-ray systems predominantly used benchmark datasets, especially the COVID-19 X-ray image database, and no system used real-time data.
- Architecture Comparison: Some reviewed systems performed better with custom networks, but results varied by dataset and were not directly comparable across different data sizes.
B. Challenges and Future Trends
The paper identifies limited, imbalanced, and heterogeneous imaging data as central obstacles for COVID-19 deep learning, and discusses transfer, ensemble, multitask, and synthetic-data strategies as possible responses.
- Challenges: Deep learning diagnosis requires large datasets, but COVID-19 lacks standard, complete, reliable, and consistently labeled imaging data.
- Challenges: Heterogeneous datasets and differing preparation and evaluation procedures make it difficult to determine which system performs best.
- Challenges: Dataset imbalance is common because COVID-19 images are fewer than pneumonia and normal cases, potentially introducing training bias.
- Challenges: The reviewed systems often lack confidence intervals for predicted values, limiting uncertainty assessment for COVID-19 diagnosis.
- Future Trends: Optimized algorithms, shallow LSTMs, and transfer-based feature extraction are suggested for settings with few COVID-19 samples.
- Future Trends: Freezing parameters, ensemble learning, and multi-task learning are proposed to reduce data demands or combine information across tasks.
- Future Trends: Data augmentation and GANs are identified as synthetic-data approaches for addressing class imbalance, although augmentation may risk overfitting.
VI. CONCLUSION
The review organizes deep-learning COVID-19 diagnosis research across CT and X-ray imaging, datasets, and model approaches, while identifying the lack of a gold standard as a major challenge. It concludes that deep learning should support, rather than replace, clinicians in diagnosis.
- The paper reviews deep-learning systems for COVID-19 diagnosis using CT and X-ray samples.
- A two-leveled taxonomy categorizes studies by deep-learning techniques and imaging modalities.
- The review compiles dataset sources to improve their accessibility and understanding for the research community.
- The lack of a gold standard is identified as the major challenge for deep-learning COVID-19 diagnosis systems.
- Deep-learning systems are presented as appropriate support tools developed with radiologists, not replacements for physicians or clinicians.