Source-linked AI summary
An Uncertainty-aware Transfer Learning-based Framework for Covid-19 Diagnosis
Afshar Shamsi Jokandan, Hamzeh Asgharnezhad, Shirin Shamsi Jokandan, Abbas Khosravi, Parham M. Kebria, Darius Nahavandi, Saeid Nahavandi, Dipti Srinivasan
TL;DR
Limited PCR availability and concerns about reliable COVID-19 diagnosis motivate an imaging-based alternative. The paper combines frozen pretrained CNN feature extraction, machine-learning classification, and epistemic-uncertainty estimation for X-ray and CT images. Linear SVM and neural-network classifiers perform best overall, while predictive uncertainty is higher for CT than X-ray images.
Problem
Limited PCR availability and concerns about diagnostic reliability motivate investigation of medical imaging for COVID-19 detection.
Method
The framework uses four pretrained CNNs to extract X-ray and CT features, machine-learning classifiers to identify cases, and ensembles to estimate epistemic uncertainty.
Results
Linear SVM and neural-network models achieve the best results across accuracy, sensitivity, specificity, and AUC, while CT predictive uncertainty is higher than X-ray uncertainty.
Takeaways & Limitations
The reported results support combining transfer learning with uncertainty estimates for COVID-19 diagnosis from medical images.
Abstract
from arXiv · showhide
The early and reliable detection of COVID-19 infected patients is essential to prevent and limit its outbreak. The PCR tests for COVID-19 detection are not available in many countries and also there are genuine concerns about their reliability and performance. Motivated by these shortcomings, this paper proposes a deep uncertainty-aware transfer learning framework for COVID-19 detection using medical images. Four popular convolutional neural networks (CNNs) including VGG16, ResNet50, DenseNet121, and InceptionResNetV2 are first applied to extract deep features from chest X-ray and computed tomography (CT) images. Extracted features are then processed by different machine learning and statistical modelling techniques to identify COVID-19 cases. We also calculate and report the epistemic uncertainty of classification results to identify regions where the trained models are not confident about their decisions (out of distribution problem). Comprehensive simulation results for X-ray and CT image datasets indicate that linear support vector machine and neural network models achieve the best results as measured by accuracy, sensitivity, specificity, and AUC. Also it is found that predictive uncertainty estimates are much higher for CT images compared to X-ray images.
I. INTRODUCTION
COVID-19 created an urgent need for early diagnosis, while limited PCR availability increased interest in accessible X-ray and CT imaging. Existing imaging models faced limited-data and epistemic-uncertainty concerns, motivating an uncertainty-aware transfer-learning framework.
- COVID-19 caused a worldwide pandemic, prompting restrictions and substantial public-health and economic impacts.
- Early screening and diagnosis are presented as important for preventing further outbreak and supporting faster medical recovery.
- X-ray and CT imaging are more accessible in many healthcare settings, while real-time PCR is unavailable at scale in many countries.
- Training deep CNNs requires substantial data and computation, but COVID-19 datasets often contain only limited numbers of labeled images.
- Limited samples raise epistemic-uncertainty and out-of-distribution concerns because models may fail on cases unlike their training data.
- The proposed framework combines pretrained CNN feature extraction, machine-learning classification, and epistemic-uncertainty estimation for X-ray and CT diagnosis.
II. PROPOSED METHOD
The framework uses frozen pretrained CNNs to extract image features, classifiers to predict COVID-19 status, and ensembles to estimate epistemic uncertainty. Prediction entropy summarizes uncertainty in the ensemble's averaged output distribution.
- Transfer learning addresses COVID-19's limited data and computational constraints by adapting pretrained ImageNet networks for medical-image classification.
- Frozen convolutional layers extract hierarchical features from X-ray and CT images, which are then passed to machine-learning classifiers.
- The framework considers aleatoric uncertainty from data noise and epistemic uncertainty from model ignorance, focusing mainly on the latter.
- An ensemble of diverse models produces predictive uncertainty estimates for new inferences.
- Prediction entropy measures epistemic uncertainty by evaluating the ensemble's mean predictive distribution.
- The ensemble averages repeated model predictions to obtain a final output probability for each class.
A. Datasets
The study uses chest X-ray and CT datasets as inputs for COVID-19 classification. The X-ray data combine COVID-19 images with non-COVID cases that may include other respiratory conditions, while the CT dataset contains 349 COVID and 397 non-COVID images.
- The study uses chest X-ray and CT imagery as information sources for COVID-19 diagnosis and applies classifiers to CNN-extracted features.
- The X-ray dataset includes 25 COVID-19 images and 75 non-COVID images collected from cited sources.
- Non-COVID X-ray cases may include bacterial or viral infections, chronic obstructive pulmonary disease, or multiple conditions rather than healthy lungs.
- The CT dataset contains 349 COVID images and 397 non-COVID images.
B. Pretrained Models
The framework compares four pretrained CNN architectures for hierarchical feature extraction. Their designs differ in connectivity and depth, while transfer learning keeps their weights frozen and uses architecture-specific input sizes.
- VGG uses a comparatively simple convolutional architecture with 3×3 filters, pooling layers, and fully connected layers.
- ResNet uses residual blocks that provide direct paths to early layers and facilitate gradient flow during backpropagation.
- DenseNet connects each layer to all preceding feature maps, reducing parameters and relieving gradient vanishing.
- InceptionResNetV2 combines inception and residual blocks to support deeper convolutional networks while addressing overfitting.
- Transfer learning freezes network weights, using 224×224 inputs for VGG16, ResNet50, and DenseNet121 and 299×299 inputs for InceptionResNetV2.
C. Classification Methods
The framework treats COVID-19 detection as binary image classification, extracting hierarchical features with pretrained networks before applying multiple classifiers. The section describes kNN and SVM-based classification principles used for these features.
- COVID-19 detection maps a chest X-ray or CT image to a binary label indicating COVID-19 presence or absence.
- Hierarchical features from five pretrained networks are processed by multiple classifiers, including kNN and linear support vector methods.
- kNN classifies samples by feature-space similarity using k = 2 and the Minkowski distance metric.The passage notes that Euclidean and Minkowski distances are commonly used.
- SVM separates classes with a hyperplane and seeks a large margin from the nearest samples to reduce generalization error.
3) RBF SVM:
The section introduces RBF-kernel classification alongside Gaussian-process, neural-network, and random-forest models for processing extracted features. These methods differ in their similarity calculations, uncertainty outputs, nonlinear mappings, and ensemble construction.
- RBF SVM: RBF SVM calculates similarity between samples with a radial basis function kernel.
- Gaussian process: Gaussian processes model joint distributions with multivariate-normal variables and provide prediction uncertainty estimates.The paper uses an RBF kernel with length-scale equal to one for its GP classifiers.
- Neural network: A feedforward neural network learns a nonlinear mapping from fixed-size inputs to targets through hidden layers and weighted neurons.The connection weights can be trained with stochastic gradient descent or Adam.
- Random forest: A random-forest classifier combines parallel decision trees trained on randomly selected feature and sample subsets, using majority voting for prediction.The paper sets the number of decision trees to 10.
7) Adaboost:
Adaboost constructs a classifier by combining weak classifiers sequentially, with each successive classifier addressing errors made earlier.
- Adaboost combines 50 weak classifiers in series, with each classifier focusing on correcting previous mistakes.
8) Naive Bayes:
Naive Bayes classifiers apply Bayes theorem, and the paper uses a Gaussian variant with a normal prior to predict posterior probabilities.
- Gaussian naive Bayes predicts a posterior using a normal prior under Bayes theory.
- The section directs readers to standard machine-learning and statistical references for fuller classifier information.
IV. SIMULATIONS AND RESULTS
The study evaluates classifiers using features extracted by four pretrained CNN architectures on CT and X-ray datasets, with repeated runs and uncertainty analysis. Linear SVM and neural-network classifiers generally perform best, while CT predictions show higher uncertainty.
- Increasing feature count reduces the classification ability of RBF SVM and GP, whereas linear SVM and neural networks remain the strongest classifiers.The authors associate the RBF SVM and GP degradation with covariance-matrix inversion and explain the stronger alternatives through linear class separation.
- Each classifier is trained and evaluated 100 times, with averaged predictions and accuracy, sensitivity, specificity, and AUC reported for CT and X-ray datasets.Reported metric values are percentages.
- 87.9% accuracy is achieved by ResNet50 with a linear SVM, exceeding the 84.7% accuracy reported for the comparison method.The authors attribute the improvement to hierarchical ResNet50 feature extraction and classifier selection.
- Linear SVM and neural-network classifiers achieve the highest AUC values among the evaluated classifiers.Their performance depends strongly on the quality of features extracted by the pretrained CNNs.
- Larger CNNs do not necessarily produce more informative or discriminative features, and CNN architecture choice has a direct impact on COVID-19 classification performance.VGG16 is identified as the most efficient pretrained network in the discussed parameter-feature analysis.
- CT images exhibit high predictive uncertainty, reflecting inconsistent labels across ensemble neural networks and motivating extra care when using CT-based predictions.Uncertainty visualizations use PCA-reduced feature spaces for X-ray and CT samples.
V. CONCLUSION
The conclusion presents deep transfer learning as a way to extract diagnostic features from medical images and identifies several directions for improving the framework. These include fine-tuning, feature or prediction combinations, and more comprehensive uncertainty estimation.
- The study investigates deep transfer learning for COVID-19 diagnosis using chest X-ray and CT images because large image repositories for training networks from scratch are unavailable.
- Fine-tuning transfer-learning models could improve the extraction of informative and discriminative features.
- Combining features from different transfer-learning models or predictions from individual models could produce hybrid models and ensembles.
- State-of-the-art methods could support more comprehensive estimation of predictive uncertainty.