COVID-19 Detection from Chest X-Ray Images Using Deep Convolutional Neural Networks with Weights Imprinting Approach

Hala As'ad, Hilda Azmi, Pengcheng Xi, Ashkan Ebadi, Stéphane Tremblay, Alexander WongJournal of Computational Vision and Imaging Systems 2021 · DOI

Abstract

The COVID-19 pandemic has drastically changed our lives. Chest radiography has been used to detect COVID-19. However, the number of publicly available COVID-19 x-ray images is extremely limited, resulting in a highly imbalanced dataset. This is a challenge when using deep learning for classification and detection. In this work, we propose the use of pre-trained deep Convolutional Neural Networks (CNN) and integrate them with a few-shot learning approach named imprinted weights. The integrated model is fine tuned to enhance the capability of detecting COVID-19. The proposed solution then combines the fine-tuned models using a weighted average ensemble for achieving an optimal 82% sensitivity to COVID-19. To the best of authors’ knowledge, the proposed solution is one of the first to utilize imprinted weights model with weighted average ensemble for enhancing the model sensitivity to COVID-19.

Introduction

Deep learning models typically require thousands of labeled examples to learn new categories effectively. During the early stages of the pandemic, while "Normal" and "Pneumonia" X-rays were plentiful, confirmed "COVID-19" images were scarce. In the COVIDx dataset, only about 3.5% of the images are COVID-positive.

This creates a significant class imbalance. When a standard model is trained on such data, it often becomes biased toward the majority classes, essentially ignoring the rare COVID-19 cases to achieve high overall accuracy. This research addresses this "needle in a haystack" problem, aiming to make models highly sensitive to rare COVID-19 cases without requiring a massive influx of new data. By using a "weights imprinting" few-shot learning technique, the researchers achieved 82% sensitivity for COVID-19 detection—drastically outperforming standard training methods that managed as little as 1% sensitivity on the same imbalanced data.

Methods

The proposed solution utilizes a multi-step pipeline combining few-shot learning with ensemble modeling:

  1. Base Training: The process begins with two architectures: a pre-trained ResNet-50 and an untrained AlexNet. These are initially trained to distinguish only between the dominant classes (Normal vs. Pneumonia).
  2. Weights Imprinting: Instead of randomly initializing the final classification layer for the new COVID-19 class, the researchers "imprint" weights. COVID-19 images are passed through the model to generate an embedding vector, which is then used to directly set the initial weights for the COVID-19 node in the final layer.
  3. Fine-Tuning: The models are refined using two strategies: k-fold cross-validation with balanced sets and random oversampling.
  4. Weighted Ensemble: Finally, four variations of these models are combined using a weighted average ensemble. A greedy search algorithm identifies the specific weights for each model that maximize sensitivity to COVID-19.
The proposed multi-model ensemble architecture utilizing weights imprinting and weighted average ensemble.

Key Insights and Results

The study utilized the COVIDx dataset, consisting of 14,198 chest X-rays where only ~3.5% were labeled as COVID-19. The core finding is that initialization matters; by using imprinted weights, the model starts fine-tuning from a knowledgeable position, preventing majority classes from overwhelming the learning process.

In medical screening, missing a positive case (a false negative) is often worse than a false alarm. Therefore, the authors specifically optimized for sensitivity. The proposed approach showed a massive improvement over baseline methods:

  • Sensitivity: The imprinted ensemble achieved 82% sensitivity for COVID-19. In comparison, a standard "All Classes Joint" model achieved only 1%, and a model trained on a small balanced subset achieved 34%.
  • Overall Accuracy: The final model reached 85% total accuracy across all three classes.
  • Precision: COVID-19 precision was 70%, significantly higher than baseline models.
Comparison of sensitivity across classes for different training approaches, highlighting the 82% sensitivity achieved by the proposed method for COVID-19.

Conclusion

While 82% sensitivity is a strong result for highly imbalanced data, the authors acknowledge that sensitivity drops as the percentage of COVID-19 data decreases, suggesting a floor to how much imprinting can compensate for data scarcity. Future work will involve comparing this approach against other few-shot methods like Triplet or Prototypical Networks and exploring advanced image pre-processing to help models focus on specific lung abnormalities associated with the virus.

Summary updated

Abstract

COVID-19 pandemic has drastically changed our lives. Chest radiographyhas been used to detect COVID-19. However, the numberof publicly available COVID-19 x-ray images is extremely limited,resulting in a highly imbalanced dataset. This is a challenge whenusing deep learning for classification and detection. In this work, wepropose the use of pre-trained deep Convolutional Neural Networks(CNN) and integrate them with a few-shot learning approach namedimprinted weights. The integrated model is fine tuned to enhancethe capability of detecting COVID-19. The proposed solution thencombines the fine-tuned models using a weighted average ensemblefor achieving an optimal 82% sensitivity to COVID-19. To thebest of authors’ knowledge, the proposed solution is one of the firstto utilize imprinted weights model with weighted average ensemblefor enhancing the model sensitivity to COVID-19.
BibTeX
@article{asad2021covid,
  author = {Hala As'ad and Hilda Azmi and Pengcheng Xi and Ashkan Ebadi and Stéphane Tremblay and Alexander Wong},
  title = {{COVID-19 Detection from Chest X-Ray Images Using Deep Convolutional Neural Networks with Weights Imprinting Approach}},
  journal = {Journal of Computational Vision and Imaging Systems},
  year = {2021},
  doi = {10.15353/jcvis.v6i1.3546},
}
Figure