Source-linked AI summary
Towards an Effective and Efficient Deep Learning Model for COVID-19 Patterns Detection in X-ray Images
Eduardo Luz, Pedro Lopes Silva, Rodrigo Silva, Ludmila Silva, Gladston Moreira, David Menotti
TL;DR
COVID-19 screening requires alternatives to conventional diagnosis, while existing approaches face computational-cost barriers. This paper exploits efficient convolutional architectures, achieving high reported screening performance with improved computational efficiency, though validation on larger and more heterogeneous databases remains necessary.
Problem
COVID-19 diagnosis involves a conventional method described as the gold standard, while computational processing time can be prohibitive for screening approaches.
Method
The paper exploits an efficient convolutional network architecture, including the EfficientNet family, for COVID-19 screening.
Results
93.9% accuracy, 96.8% COVID-19 sensitivity, and 100% positivity prediction were reported with computational efficiency improvements over baseline work.
Takeaways & Limitations
The proposed approach produced a high-quality model while improving computational efficiency compared with baseline work.
Takeaways & Limitations
Larger and more heterogeneous databases are still needed to validate the methods before claiming that deep learning can assist physicians in detecting COVID-19 in X-ray images.
Abstract
from arXiv · showhide
Confronting the pandemic of COVID-19, is nowadays one of the most prominent challenges of the human species. A key factor in slowing down the virus propagation is the rapid diagnosis and isolation of infected patients. The standard method for COVID-19 identification, the Reverse transcription polymerase chain reaction method, is time-consuming and in short supply due to the pandemic. Thus, researchers have been looking for alternative screening methods and deep learning applied to chest X-rays of patients has been showing promising results. Despite their success, the computational cost of these methods remains high, which imposes difficulties to their accessibility and availability. Thus, the main goal of this work is to propose an accurate yet efficient method in terms of memory and processing time for the problem of COVID-19 screening in chest X-rays. Methods: To achieve the defined objective we exploit and extend the EfficientNet family of deep artificial neural networks which are known for their high accuracy and low footprints in other applications. We also exploit the underlying taxonomy of the problem with a hierarchical classifier. A dataset of 13,569 X-ray images divided into healthy, non-COVID-19 pneumonia, and COVID-19 patients is used to train the proposed approaches and other 5 competing architectures. Finally, 231 images of the three classes were used to assess the quality of the methods. Results: The results show that the proposed approach was able to produce a high-quality model, with an overall accuracy of 93.9%, COVID-19, sensitivity of 96.8% and positive prediction of 100%, while having from 5 to 30 times fewer parameters than other than the other tested architectures. Larger and more heterogeneous databases are still needed for validation before claiming that deep learning can assist physicians in the task of detecting COVID-19 in X-ray images.
1. Introduction
The paper targets COVID-19 screening from chest X-rays with deep learning, emphasizing both detection quality and computational efficiency. It investigates EfficientNet-based flat and hierarchical classifiers for models suitable for constrained devices.
- Motivation: COVID-19 screening is motivated by the time and supply limitations of RT-PCR testing.The paper presents chest X-rays and deep learning as alternative approaches for detecting COVID-19 patterns.
- Motivation: Deep learning is considered for chest-radiography analysis because convolutional neural networks have performed strongly in computer-vision tasks.Earlier work had also explored deep learning for pneumonia and other diseases on radiographs.
- Objective: The work seeks models that detect chest-X-ray patterns, including patterns imperceptible to the human eye, while reducing computational cost.Lower cost is linked to using higher-resolution images without prohibitive processing time and to deployment in smartphones or medical equipment.
- Implications: The work may support future medical applications using a camera on a regular cellphone, with source code and pretrained models made available.The introduction frames low memory use and low latency as requirements for embedded devices and large-scale deployment.
- Approach: EfficientNet architectures and hierarchical classification are investigated to obtain cost-efficient models and exploit the problem’s class taxonomy.The hierarchical approach requires multiple models, making computational cost especially relevant.
- Results: The proposed models are reported as smaller without compromising accuracy, making fast mobile-device inference more feasible.The paper also states that the hierarchical model uses more computational resources and is less effective for minority classes, including COVID-19.
- Results: 93.9% overall accuracy, 96.8% COVID-19 sensitivity, and 100% positivity prediction were achieved by the proposed model.The reported positivity prediction corresponds to no false positives.
2. Related Works
Related studies applied convolutional and pretrained neural networks, feature extraction, and conventional classifiers to COVID-19 chest-X-ray classification. Their reported results vary across datasets, class definitions, and evaluation settings, limiting direct comparison.
- Evaluation differences: Another study reported an F1-score of 0.89 for COVID-19 using a dataset of only 50 images.The reviewed paper states that direct comparison is not possible because the datasets differ in nature.
- Deep-learning architectures: Prior studies evaluated CNN architectures including COVID-net, ResNet50, CoroNet, VGG19, and DenseNet201 for chest-X-ray classification.The reviewed tasks used different numbers and types of classes, including normal, COVID-19, and bacterial or viral pneumonia.
- Reported results: CoroNet achieved 89.6% accuracy, with 93% precision and 98.2% recall for COVID-19 cases in a 4-class setting.It used an Xception CNN pretrained on ImageNet and was evaluated on databases assembled from two public image sources.
- Evaluation differences: Some evaluation protocols were potentially compromised when a second test dataset apparently contained COVID-19 images used during training.This issue affects interpretation of the corresponding reported results.
3. Problem Setting
The task is to classify a chest X-ray as normal, COVID-19, or non-COVID-19 pneumonia. The paper considers both flat classification and hierarchical classification based on a taxonomy of the target classes.
- Target task: Given a chest X-ray, the model must distinguish healthy patients, COVID-19 patients, and patients with other forms of pneumonia.These three labels are named normal, COVID-19, and pneumonia.
- Target task: The model should not assume the view in which the chest X-ray was taken.The task is defined over images similar to typical samples from the COVIDx dataset.
- Clinical motivation: The three-class output may help clinicians prioritize patients for PCR confirmation and may inform treatment selection.The paper notes that COVID-19 and non-COVID-19 infections require different treatment plans.
- Classification formulations: The study compares flat classification, which disregards class relationships, with hierarchical classification, which assumes the classes form a taxonomy.In hierarchical classification, target classes occupy the leaves of the taxonomy.
4. Methodology
The methodology combines COVIDx data construction, EfficientNet-based architectures, and compound scaling to address chest-X-ray classification. The study adapts a segmentation dataset for classification and uses EfficientNet’s compact architecture as its baseline.
- Datasets: The methodology uses chest X-rays to classify Normal, Pneumonia, and COVID-19 cases.The RSNA dataset supplies Normal and Pneumonia images, while the COVID-19 Image Data Collection supplies COVID-19 cases.
- Datasets: COVIDx combines the RSNA Pneumonia Detection Challenge dataset with the COVID-19 Image Data Collection.It contains 13,800 images from 13,645 individuals and is divided into training and testing partitions.
- EfficientNet: The EfficientNet baseline uses the Mobile Inverted Bottleneck Convolution block as its main component.The MBConv block includes depthwise convolution, batch normalization, tensor dimensions, and repeated-layer multipliers.
- EfficientNet: EfficientNet starts from a compact baseline and systematically scales depth, width, and resolution using fixed coefficients.The coefficient φ controls available resources, while the scaling equation determines changes in model FLOPS.
- EfficientNet: EfficientNet achieved 8.4x fewer parameters and was 6.1x faster than GPipe on ImageNet.This prior result motivates EfficientNet as a high-performance architecture with lower computational cost.
4.3. Hierarchical Classification
The paper uses hierarchical classification to reflect the relationships among Normal, Pneumonia, and COVID-19 classes. Separate classifiers operate at successive levels of the class tree.
- Classification schemes: Local classification uses a set of classifiers for each tree level, whereas flat classification ignores the hierarchy.Global classification instead uses one classifier able to classify any node in the tree.
- Hierarchical classification: Hierarchical classification places target classes at tree leaves and classifiers at intermediate nodes.The paper identifies this local classification strategy as its proposed approach.
- Two-level hierarchy: The hierarchy first separates Normal from Pneumonia, then distinguishes pneumonia types.The problem is represented with two classification levels.
- Inference procedure: Inference ends after the root classifier predicts Normal; Pneumonia cases proceed to a second classifier.The second classifier determines the relevant pneumonia type.
4.4. Training
Training combines preprocessing, augmentation, transfer learning, and EfficientNet extensions. The proposed models add four blocks and use regularization and swish activation in the added layers.
- Image preprocessing: The images receive intensity normalization to [0, 1], while convolutional filters perform possible data cleaning.Images are also resized according to the architecture’s resolution parameter.
- Data augmentation: Data augmentation applies rotation, horizontal flipping, and scaling without hindering radiographic interpretation.The specified transformations include rotation between 0 and 15 degrees and 20% zoom.
- Proposed models: The proposed models extend EfficientNet by adding four operator blocks atop the base network.EfficientNet B0 is used as the base model for the proposed architectures.
- Proposed models: The added classification layers use fully connected layers, batch normalization, dropout, and swish activation.Batch normalization regularizes and stabilizes training, while dropout inhibits randomly selected neurons during training.
- Transfer learning: Transfer learning initializes EfficientNet with ImageNet weights and fine-tunes all layers on the new training phase.Training updates the weights with Adam and reduces the learning rate by a factor of 10 after stagnation.
4.5. Model evaluation and metrics
The final evaluation uses the COVIDx dataset and assesses classification with accuracy, COVID-19 sensitivity, and COVID-19 positive prediction. Multiply-accumulate operations measure computational cost.
- Evaluation dataset: The final evaluation is carried out with the COVIDx dataset.COVIDx combines two public datasets, and the authors follow a published script to load its training and test sets.
- Metrics: The evaluation reports accuracy, COVID-19 sensitivity, and COVID-19 positive prediction.These metrics are denoted Acc, SeC, and +PC, respectively.
- Metrics: The confusion-matrix terms distinguish correctly classified Normal, non-COVID-19, and COVID-19 samples from COVID-19 misclassifications.The definitions include COVID-19 samples classified as Normal or non-COVID-19.
- Efficiency: Multiply-accumulate operations measure the computational cost of the models.This complements the predictive metrics with an efficiency measure.
5. Experiments and Discussion
Experiments compare EfficientNet-based flat and hierarchical classifiers across training configurations and network architectures. Balanced flat classification produced the strongest results, while the proposed approach reduced computational demands and retained competitive diagnostic performance.
- Experimental setup: Three training configurations were evaluated: Raw, Raw + Data Augmentation, and Balanced datasets.The balanced dataset used 1,000 images per class through COVID-19 augmentation and undersampling of the other classes.
- Experimental setup: The hierarchical setup used two classifiers: Normal versus Pneumonia, followed by COVID-19 versus other pneumonia.COVID-19 and common Pneumonia were combined for the first classification level, then separated at the second level.
- Classification results: Balanced flat classification achieved better results than the evaluated hierarchical configuration.The balanced flat scenario was selected for comparisons among the base network architectures.
- Classification results: All tested networks had comparable accuracy, but more complex models performed worse on the COVID-19 minority class.The authors hypothesize that the limited dataset may explain the difficulty of fitting more parameterized models.
- Efficiency comparison: Compared with COVID-Net, the proposed approach improved positive prediction without losing sensitivity while reducing memory and latency.The proposed model was reported as more than 15 times smaller under a comparable evaluation protocol.
- Model interpretation: Activation maps highlighted opaque regions in correctly classified COVID-19 images, whereas healthy-image maps highlighted non-opaque regions.The displayed examples came from the test set and were not seen during training.
- Hierarchical analysis: The hierarchical model classified the Normal class better but reduced COVID-19 sensitivity and positive prediction.The authors associate this trade-off with shared features between pneumonia classes and reduced training diversity after removing Normal images.
6. Findings and Future Direction
The authors summarize an efficient COVID-19 detection approach intended for deployment on conventional smartphones and medical equipment. They report strong screening metrics and discuss hierarchical classification as potentially less affected by evaluation-protocol bias, while calling for broader validation data.
- Main findings: An efficient, low-computational approach was proposed to detect COVID-19 from chest X-ray images.The stated deployment direction includes medical equipment and physicians’ mobile phones.
- Main findings: 90% sensitivity and 100% positive prediction were obtained despite few COVID-19 images in the evaluation protocol.The authors describe these as insightful results under the stated protocol.
- Future direction: The hierarchical approach is reported to suffer less from bias in the evaluation protocols discussed by the authors.This is presented as a significant gain that supports using hierarchical analysis for the task.
- Limitations and reproducibility: Larger and more heterogeneous databases are needed for validation, while additional COVID-19 images may improve training and make testing more challenging.The authors also provide code to support reproducibility and future comparisons.
- Deployment direction: The proposed model requires 55Mb of memory and has viable inference time on a conventional cell-phone processor.The approach is presented as compatible with edge computing without relying on internet or cloud services.
7. Conclusion
The paper evaluates efficient convolutional approaches, including flat and hierarchical classification, for detecting COVID-19-related abnormalities in chest radiographs. The proposed approach achieved strong reported performance and computational efficiency, but broader validation remains necessary.
- 7. Conclusion: The experiments evaluated flat and hierarchical classification approaches on the COVIDx dataset.Transfer learning and data augmentation supported effective training despite limited COVID-19 image counts.
- 7. Conclusion: More than 30 times higher computational efficiency was reported compared with baseline work.The proposed approach was presented as an efficient, low-computational method for COVID-19 detection.
- 7. Conclusion: The approach was described as a promising candidate for embedding in medical equipment or physicians’ mobile phones.This proposed deployment scope follows the reported efficiency and screening performance.
- 7. Conclusion: Larger and more heterogeneous databases are still needed to validate the methods before claiming clinical assistance from deep learning.The conclusion also notes that current datasets contain limited numbers of COVID-19-related images.
Conflict of interest statement
The authors report no known competing financial interests or personal relationships that could have influenced the work.
- Conflict of interest statement: The authors declare no known competing financial interests.
- Conflict of interest statement: The statement addresses both financial and personal sources of potential influence.
- Conflict of interest statement: The authors declare no personal relationships that could have influenced the reported work.