Source-linked AI summary
Extracting possibly representative COVID-19 Biomarkers from X-Ray images with Deep Learning approach and image data related to Pulmonary Diseases
Ioannis D. Apostolopoulos, Sokratis Aznaouridis, Mpesiana Tzani
TL;DR
The study addresses the need for automatic COVID-19 and pulmonary-disease classification from X-ray images amid limited large-scale data. It evaluates MobileNet v2 feature-extraction strategies, finding approximately 99% COVID-19 diagnostic accuracy and suggesting extracted features may warrant biomarker analysis.
Problem
Automatic, reliable classification of COVID-19 and other pulmonary diseases from X-ray images is needed, while large-scale image data remain limited.
Method
The study evaluates MobileNet v2 for distinguishing six common pulmonary diseases, comparing training from scratch with off-the-shelf feature extraction and fine-tuning.
Results
Approximately 99% accuracy in COVID-19 diagnosis was achieved, while transfer-learning strategies underperformed in the particular experiment.
Takeaways & Limitations
The findings suggest extracted features may be biomarkers and that automatic COVID-19 detection can reduce medical staff exposure to the outbreak.
Takeaways & Limitations
The dataset includes a relatively small COVID-19 sample that may overrepresent severe cases, with mild cases missing.
Abstract
from arXiv · showhide
In this study, the problem of automatically classifying pulmonary diseases, including the recently emerged COVID-19, from X-Ray images, is considered. While the spread of COVID-19 is increased, new, automatic, and reliable methods for accurate detection are essential to reduce the exposure of the medical experts to the outbreak. X-ray imaging, although limited to specific visualizations, may be helpful for the diagnosis. Deep Learning has proven to be a remarkable method to extract massive high-dimensional features from medical images. Specifically, in this paper, the state-of-the-art Convolutional Neural Network called Mobile Net is employed and trained from scratch to investigate the importance of the extracted features for the classification task. A large-scale dataset of 3905 X-Ray images, corresponding to 6 diseases is utilized for training MobileNet v2, which has been proven to achieve remarkable results in related tasks. The results suggest that training CNNs from scratch may reveal vital biomarkers related but not limited to the COVID-19 disease, while an overall classification accuracy of the seven classes reaches 87.66%. Besides, this method achieves 99.18% accuracy, 97.36% Sensitivity, and 99.42% Specificity in the detection of COVID-19.
1 Introduction
The study addresses automated pulmonary-disease classification from X-Ray images, motivated by COVID-19 detection needs and limited large-scale data. It evaluates MobileNet-based feature extraction, emphasizing training from scratch and possible biomarker discovery.
- COVID-19 detection from thoracic X-Ray scans is presented as a need for protecting healthcare workers and supporting diagnosis, quarantine, and follow-up.
- Existing COVID-19 X-Ray studies are limited by incomplete data and often classify only viral or bacterial pneumonia, COVID-19, and normal images.
- MobileNet v2 is evaluated across training from scratch, transfer learning, and fine-tuning to examine extracted features across six common pulmonary diseases including COVID-19.
- The experiment combines common pulmonary-disease X-Rays with confirmed COVID-19 cases because a complete dataset containing both was unavailable.
- 88% classification accuracy and approximately 99% COVID-19 diagnostic accuracy are reported, with the extracted features proposed for further biomarker analysis.
2.1 Dataset of the study
The study constructs a multi-disease X-Ray dataset by combining confirmed COVID-19 images, pneumonia images, and selected pulmonary abnormalities. The authors identify labeling, sampling, clinical-context, and demographic limitations that constrain interpretation.
- Dataset construction: The COVID-19 collection contains 455 confirmed pathological X-Ray images gathered from public repositories.
- Dataset construction: 910 bacterial and viral pneumonia X-Ray images were added to evaluate classification under more complex pulmonary-disease conditions.
- Dataset limitations: NIH images use disease labels assigned by automatic text mining of radiological reports, with estimated NLP labeling accuracy above 90%.
- Dataset construction: The selected additional diseases are pulmonary edema, pleural effusion, COPD, and pulmonary fibrosis, chosen for significance and frequency.
- Dataset limitations: The COVID-19 sample is relatively small and may overrepresent severe cases because mild-symptom cases are absent from public collections.
- Dataset limitations: Older pneumonia samples lack suspected-Coronavirus clinical conditions, while demographic and other risk-factor data are unavailable.
2.2 Learning Strategies for Feature Extraction
The paper compares training from scratch with transfer-learning approaches for extracting CNN features. Fine-tuning partially retrains pretrained networks, with progressively more convolutional blocks made trainable.
- The study compares training from scratch, pretrained feature extraction, and fine-tuning as three CNN feature-extraction strategies.
- Transfer learning reuses a model developed for one task on another task and includes off-the-shelf extraction and fine-tuning.
- Off-the-shelf extraction uses source-task convolutional weights without retraining the network and adds a classifier for the target task.
- Fine-tuning initializes the network with pretrained weights and partially retrains it on the target task.
- The experiment defines six fine-tuning cases that progressively increase the number of trainable convolutional blocks.
2.3 Method for the extraction of possibly significant biomarkers
The study uses MobileNet v2 to extract image features for pulmonary-disease classification, emphasizing training from scratch to assess whether image-specific features may represent biomarkers.
- MobileNet was selected for classification because it uses fewer parameters and supports swift training.The model was also reported to reduce false negatives for COVID-19 detection relative to other CNNs.
- MobileNet reduces parameters through depthwise separable convolutions that combine depthwise and pointwise convolutions.
- The architecture adds global average pooling, a 2500-node neural network, batch normalization, and dropout to process extracted features and reduce overfitting.Dropout randomly zeros 50% of neuron outputs.
- Training from scratch uses the MobileNet architecture without learned parameters from pre-trained models, enabling evaluation of features extracted from the study’s X-ray images.The authors contrast this with transfer learning based on features learned from non-medical images.
- The extracted features may constitute real biomarkers related to various diseases.
2.4 Experiment Setup
The experiment compares three learning strategies using the same MobileNet v2, with evaluation performed through 10-fold cross-validation.
- Three experiments use the same MobileNet v2 while varying the learning strategy.
- The evaluated strategies are transfer learning with off-the-shelf features, transfer learning with fine-tuning, and training from scratch.Training from scratch borrows the architecture but not the learned parameters.
- Training and evaluation use 10-fold cross-validation, rotating each fold as the hidden test set.Nine folds train the model while one fold tests performance and prediction confidence.
2.5 Metrics
Performance is assessed for seven-class disease classification and binary COVID-19 detection, with sensitivity and specificity derived from prediction counts.
- The evaluation reports overall seven-class accuracy and binary accuracy for COVID-19 versus non-COVID-19 classification.
- COVID-19 detection records true positives, true negatives, false positives, and false negatives to calculate sensitivity and specificity.
- Sensitivity is calculated as true positives divided by true positives plus false negatives.
- Specificity is calculated as true negatives divided by true negatives plus false positives.
- Because of class imbalance, seven-class accuracy is treated as the most reliable metric, followed by specificity when accuracy is high.Specificity is emphasized because correctly identifying actual non-COVID-19 cases is important.
3 Results
The results compare of-the-self features, fine-tuning, and training from scratch across pulmonary-disease classification tasks. Training from scratch is reported as the optimal strategy, with strong COVID-19 discrimination and seven-class performance, while of-the-self features show substantial bias toward non-COVID-19 cases.
- Of-the-Self-features strategy: Of-the-self features achieved 88.81% 2-class accuracy, 51.98% 7-class accuracy, and 04.62% Sensitivity; Specificity was omitted because class imbalance pushed it toward 100%.The authors state that the resulting Specificity was not meaningful under the class imbalance.
- Of-the-Self-features strategy: The of-the-self-features strategy was judged ineffective for extracting useful COVID-19-related features and showed bias toward non-COVID-19 cases.This conclusion is based on the corresponding confusion-matrix analysis.
- Fine-Tuning strategy: Fine-tuning results are reported for multiple cases defined by the number of trainable CNN blocks, with accuracy calculated as the mean across 10-fold cross-validation folds.Table 4 reports Accuracy, Sensitivity, and Specificity for the different fine-tuning cases.
- Training-from-scratch strategy: In the seven-class confusion matrix, 443 of 455 COVID-19 cases were correctly identified, while only 2 were classified as normal.Among 1,341 normal cases, only 1 was mistakenly identified; other abnormalities showed more diverse classification outcomes.
- Training-from-scratch strategy: The COVID-19-versus-all-classes confusion matrix indicates an excellent trade-off among True Positives, False Positives, True Negatives, and False Negatives.The authors use this comparison to characterize the binary detection performance.
4 Discussion
The study explores whether deep-learning features from X-Ray images can serve as possible COVID-19-related biomarkers, while identifying dataset and disease-visibility boundaries that constrain interpretation.
- Deep Learning extracts millions of image features, but most may be irrelevant to the desired classification outcome.The paper distinguishes extracted image features from biomarkers and notes that CNN classification rejects many features after convolutional processing.
- High classification accuracy is presented as evidence that X-Ray images may contain reliable biomarkers, although this possibility requires further investigation.The authors suggest exploring other approaches, including Radiomics, in future research.
- Training from scratch may mine pulmonary-abnormality features relevant to biomarkers, whereas transfer learning strategies underperformed in this experiment.The paper attributes differences in extracted features to the initial training used by each strategy, while noting that transfer-learning methods perform well in other medical-image tasks.
- A more in-depth analysis requires substantially more patient data, particularly additional COVID-19 cases.
- Mild COVID-19 symptoms may not be visualized correctly, or at all, on X-rays, limiting detection of patients with less severe disease.
- Models should distinguish a larger number of pulmonary diseases, while X-Ray screening should not determine treatment by itself.The authors frame screening as useful for timely quarantine pending complete examination, specific treatment, or follow-up.
5 Conclusion
The work reports low-cost, rapid, automatic COVID-19 detection from X-Ray images and argues that extracted features warrant investigation as possible biomarkers.
- Low-cost, rapid, automatic COVID-19 detection was achieved using a significantly large sample of pulmonary infections.
- The study recommends future research into whether extracted features behave as biomarkers, based on evidence from the reported results.
- Automatic COVID-19 detection from medical images may reduce nursing and medical staff exposure to the outbreak.