[Paper Review] Transfer learning approach to Classify the X-ray image that corresponds to corona disease Using ResNet50 pretrained by ChexNet

Mahyar Bolhassani|arXiv (Cornell University)|May 18, 2021

COVID-19 diagnosis using AI8 references4 citations

TL;DR

This study proposes a transfer learning approach using ResNet50 and DenseNet121 pretrained on ImageNet and CheXNet to classify COVID-19 from chest X-rays on an imbalanced dataset. By leveraging focal loss and data augmentation, the model achieves 84.6% test accuracy after adding extra COVID-19 samples, with DenseNet121 fine-tuned on CheXNet outperforming other architectures.

ABSTRACT

Coronavirus adversely has affected people worldwide. There are common symptoms between the Covid19 virus disease and other respiratory diseases like pneumonia or Influenza. Therefore, diagnosing it fast is crucial not only to save patients but also to prevent it from spreading. One of the most reliant methods of diagnosis is through X-ray images of a lung. With the help of deep learning approaches, we can teach the deep model to learn the condition of an affected lung. Therefore, it can classify the new sample as if it is a Covid19 infected patient or not. In this project, we train a deep model based on ResNet50 pretrained by ImageNet dataset and CheXNet dataset. Based on the imbalanced CoronaHack Chest X-Ray dataset introducing by Kaggle we applied both binary and multi-class classification. Also, we compare the results when using Focal loss and Cross entropy loss.

Motivation & Objective

• To address the challenge of classifying COVID-19 in chest X-rays using deep learning on highly imbalanced datasets.
• To evaluate the effectiveness of transfer learning using ImageNet- and CheXNet-pretrained models for medical image classification.
• To compare the impact of different loss functions—cross-entropy and focal loss—on model performance under data imbalance.
• To improve model generalization by augmenting the limited number of COVID-19 X-ray samples through data synthesis and external data addition.

Proposed method

• Fine-tuned ResNet50 and DenseNet121 architectures pre-trained on ImageNet and CheXNet, respectively, for transfer learning on the CoronaHack Chest X-Ray dataset.
• Applied data augmentation techniques including random horizontal/vertical flips and random rotations to increase training sample diversity.
• Used focal loss to mitigate the adverse effects of class imbalance, particularly for rare classes like COVID-19.
• Trained models using Adam optimizer with a cosine decay learning rate schedule and early stopping to prevent overfitting.
• Generated synthetic COVID-19 X-ray images using conditional GANs, though limited by GPU constraints and hyperparameter sensitivity.
• Enhanced the dataset by adding external COVID-19 X-ray images from another Kaggle challenge to improve minority class representation.

Experimental results

Research questions

• RQ1 Does transfer learning using CheXNet-pretrained models improve classification performance on imbalanced chest X-ray datasets compared to ImageNet-pretrained or randomly initialized models?
• RQ2 How does focal loss compare to categorical cross-entropy loss in improving classification accuracy for minority classes like COVID-19?
• RQ3 To what extent does data augmentation and synthetic data generation enhance model generalization on limited real-world medical imaging data?
• RQ4 Can external COVID-19 X-ray images from public datasets significantly improve model performance when added to an existing imbalanced dataset?

Key findings

• DenseNet121 pretrained on CheXNet achieved the highest test accuracy of 84.6% after adding extra COVID-19 samples, outperforming other models.
• Focal loss improved validation and test accuracy compared to standard cross-entropy loss, particularly for the underrepresented COVID-19 class.
• The model trained on CheXNet-pretrained DenseNet121 achieved 81.83% validation accuracy and 84.6% test accuracy after data extension, indicating strong generalization.
• Even with data augmentation and focal loss, the baseline ResNet50 model without transfer learning achieved only 71.11% test accuracy, highlighting the importance of pretraining.
• The addition of external COVID-19 X-ray images significantly improved model performance, increasing test accuracy from 75.56% to 84.6%.
• Despite efforts, GAN-based data generation failed to produce realistic COVID-19 images due to training instability and hardware limitations.