[Paper Review] Deep transfer learning for detecting Covid-19, Pneumonia and Tuberculosis using CXR images -- A Review

Irad Mwendo, Kinyua Gikunda+1 more|arXiv (Cornell University)|Mar 26, 2023

COVID-19 diagnosis using AI4 citations

TL;DR

This review evaluates deep transfer learning models for detecting COVID-19, pneumonia, and tuberculosis using chest X-ray (CXR) images, emphasizing architectures like ResNet-50, VGG-16, and DenseNet-121. It highlights that data augmentation, transfer learning, and hyperparameter tuning improve accuracy, with models achieving up to 99.7% accuracy, while identifying data scarcity, class imbalance, and single-disease focus as key limitations.

ABSTRACT

Chest X-rays remains to be the most common imaging modality used to diagnose lung diseases. However, they necessitate the interpretation of experts (radiologists and pulmonologists), who are few. This review paper investigates the use of deep transfer learning techniques to detect COVID-19, pneumonia, and tuberculosis in chest X-ray (CXR) images. It provides an overview of current state-of-the-art CXR image classification techniques and discusses the challenges and opportunities in applying transfer learning to this domain. The paper provides a thorough examination of recent research studies that used deep transfer learning algorithms for COVID-19, pneumonia, and tuberculosis detection, highlighting the advantages and disadvantages of these approaches. Finally, the review paper discusses future research directions in the field of deep transfer learning for CXR image classification, as well as the potential for these techniques to aid in the diagnosis and treatment of lung diseases.

Motivation & Objective

• To evaluate the effectiveness of deep transfer learning in classifying chest X-ray images for COVID-19, pneumonia, and tuberculosis.
• To identify key deep learning architectures and techniques that enhance diagnostic accuracy in low-resource and high-workload clinical settings.
• To analyze challenges such as small datasets, data imbalance, and limited multi-disease detection in existing studies.
• To highlight the role of data augmentation and hyperparameter tuning in improving model generalization and performance.
• To outline future research directions for scalable, accurate, and robust AI-assisted diagnosis of lung diseases using CXR imaging.

Proposed method

• Systematic review of 58 recent studies using deep transfer learning for CXR-based lung disease classification.
• Evaluation of pre-trained models including ResNet-50, VGG-16, DenseNet-121, InceptionV3, and EfficientNet for transfer learning on CXR datasets.
• Application of data augmentation techniques such as flipping, rotation, zooming, and histogram equalization to increase training data diversity.
• Use of hyperparameter tuning to optimize learning rates, batch sizes, and network depth for improved model convergence and accuracy.
• Incorporation of advanced techniques like anisotropic diffusion filtering, unsharp masking, and spatial transformer networks to reduce image noise and enhance features.
• Comparison of ensemble methods (e.g., random forest, SVM, logistic regression) and hybrid models combining CNNs with attention or feature extraction modules.

Experimental results

Research questions

• RQ1 Which deep transfer learning architectures demonstrate the highest accuracy in detecting COVID-19, pneumonia, and tuberculosis from CXR images?
• RQ2 How do data augmentation and hyperparameter tuning impact the performance and robustness of transfer learning models in lung disease classification?
• RQ3 What are the primary challenges—such as data scarcity, class imbalance, and model bias—that limit the real-world deployment of these models?
• RQ4 To what extent do hybrid or ensemble models improve diagnostic accuracy compared to standalone deep learning models?
• RQ5 What are the key research gaps and future directions for developing scalable, multi-disease detection systems using CXR imaging?

Key findings

• Transfer learning models such as ResNet-50 and DenseNet-121 achieved high accuracy, with one study reporting 99.9% on a small dataset of 1,824 CXR images.
• The use of data augmentation techniques like rotation, flipping, and zooming significantly improved model generalization and training set size.
• Hyperparameter tuning, including learning rate adjustment and dropout layers, enhanced model performance, with one model achieving 94% accuracy after 100 training epochs.
• Hybrid models combining CNNs with spatial transformer networks and demographic features (age, gender, view position) achieved 73% validation accuracy despite noisy data.
• Ensemble models combining random forest, SVM, and logistic regression achieved 98% accuracy, while an ensemble deep transfer learning model reached 99.70% accuracy for obstructive pulmonary diseases.
• Challenges such as data imbalance and small dataset sizes remain prevalent, with one study reporting 85.68% accuracy after using shuffle sampling to address class imbalance in TB classification.