Deep Learning for Automated Recognition of Covid-19 from Chest X-ray Images

Linh T. Duong, Phuong T. Nguyen, Ludovico Iovino, Michele FlamminimedRxiv 2020 · DOI

Abstract

The pandemic caused by coronavirus in recent months is having a devastating global effect, putting the world under an unprecedented emergency. Since effective antiviral treatments for Covid-19 are not yet available, early detection and monitoring of disease progression are crucial to reducing mortality. While vaccines and other measures are deployed, medical imaging techniques have been investigated to assist in diagnosis. This paper presents a practical solution for detecting Covid-19 from chest X-ray (CXR) images using cutting-edge Machine Learning techniques. We employ EfficientNet and MixNet families of deep neural networks as the main classification engine, combined with advanced transfer learning strategies. Validated on two real datasets (up to 17,905 images), our approach consistently yields accuracy above 95.0%, peaking at 96.64%. These results represent a significant performance improvement over existing studies.

Introduction

Early in the COVID-19 pandemic, rapid and accurate diagnosis was a primary bottleneck for healthcare systems. While RT-PCR (swab) tests remain the gold standard, they can be slow and occasionally yield false negatives. Chest X-rays (CXR) offer a faster, more accessible alternative for monitoring lung involvement. However, the challenge for artificial intelligence lies in "differential diagnosis." COVID-19, bacterial pneumonia, and other viral pneumonias often present similar radiological features, such as bilateral lesions. Furthermore, because COVID-19 was a new disease at the time of this study, researchers faced a massive data imbalance: thousands of "Normal" and "Pneumonia" images were available, but very few confirmed "COVID-19" X-rays existed to train robust models.

Methodology

To address these challenges, the researchers moved beyond standard convolutional neural networks (CNNs) to utilize two state-of-the-art model families:

  • EfficientNet: This architecture employs a "compound scaling" method that balances network depth, width, and resolution. This allows for higher accuracy with significantly fewer parameters than traditional models like ResNet.
  • MixNet: This architecture utilizes "Mixed Depthwise Convolutional Kernels," incorporating different filter sizes (e.g., 3x3, 5x5, 7x7) within the same layer to capture both fine-grained details and larger radiological patterns simultaneously.
The deep learning architecture used for multi-class classification of chest X-rays into Covid-19, Normal, and Pneumonia categories.

Transfer Learning and Data Strategies

To mitigate data scarcity, the authors employed Transfer Learning, starting with models pre-trained on the ImageNet dataset. They evaluated three advanced training strategies:

  1. Standard ImageNet: Basic pre-training.
  2. AdvProp (Adversarial Propagation): A method using adversarial examples during training to improve model robustness and generalization to real-world clinical images.
  3. Noisy Student (NS): A semi-supervised approach where a "teacher" model labels unlabeled data, and a "student" model is trained on the combined set with added noise (dropout/augmentation) to force the learning of more robust features.

The study utilized two primary datasets: $D_1$ (15,000 images, 108 COVID-19) and $D_2$ (17,905 images, 327 COVID-19). These datasets were used to validate if performance remained stable as more COVID-19 data became available.

Results and Key Insights

The experimental results demonstrated that architecture choice and training strategy are critical for medical imaging tasks:

  • Optimal Performance: EfficientNet-B3 emerged as the top performer, achieving a maximum accuracy of 96.64% when paired with AdvProp on dataset $D_1$.
  • Consistency and Speed: Across all ten experimental configurations, accuracy never dropped below 95%. Furthermore, the models proved highly efficient; EfficientNet-B0 processed 138 images per second, while even the larger MixNet-XL managed 83 images per second, making them suitable for real-time clinical use.
  • The Power of AdvProp: For the EfficientNet family, weights pre-trained with AdvProp consistently outperformed other methods. This suggests that the robustness gained from adversarial training is particularly effective at handling the subtle variations in X-ray quality found in clinical settings.
  • Recall Challenges: While the models achieved near-perfect Precision (minimizing false positives), the Recall for COVID-19 was lower (approximately 0.70 in some tests). This indicates that the models still missed some COVID cases, likely due to the extreme class imbalance in the training data.

Conclusion and Future Work

This study demonstrates that EfficientNet and MixNet architectures, combined with advanced transfer learning, can provide highly accurate automated screening for COVID-19. The authors even developed a prototype mobile application to show the potential for point-of-care assistance.

However, the researchers acknowledge limitations regarding "Internal Validity" due to the relatively small COVID-19 sample size. Future work will focus on incorporating larger, more diverse datasets to improve COVID-19 recall and fine-tuning MixNet hyperparameters. While promising, the authors emphasize that these models are currently research prototypes and require further clinical validation before deployment in medical guidance.

Summary updated

Abstract

Abstract Background The pandemic caused by coronavirus in recent months is having a devastating global effect, which puts the world under the most ever unprecedented emergency. Currently, since there are not effective antiviral treatments for Covid-19 yet, it is crucial to early detect and monitor the progression of the disease, thus helping to reduce mortality. While a corresponding vaccine is being developed, and different measures are being used to combat the virus, medical imaging techniques have also been investigated to assist doctors in diagnosing this disease. Objective This paper presents a practical solution for the detection of Covid-19 from chest X-ray (CXR) images, exploiting cutting-edge Machine Learning techniques. Methods We employ EfficientNet and MixNet, two recently developed families of deep neural networks, as the main classification engine. Furthermore, we also apply different transfer learning strategies, aiming at making the training process more accurate and efficient. The proposed approach has been validated by means of two real datasets, the former consists of 13,511 training images and 1,489 testing images, the latter has 14,324 and 3,581 images for training and testing, respectively. Results The results are promising: by all the experimental configurations considered in the evaluation, our approach always yields an accuracy larger than 95.0%, with the maximum accuracy obtained being 96.64%. Conclusions As a comparison with various existing studies, we can thus conclude that our performance improvement is significant.
BibTeX
@misc{duong2020deep,
  author = {Linh T. Duong and Phuong T. Nguyen and Ludovico Iovino and Michele Flammini},
  title = {{Deep Learning for Automated Recognition of Covid-19 from Chest X-ray Images}},
  year = {2020},
  doi = {10.1101/2020.08.13.20173997},
  howpublished = {medRxiv},
}
Figure