Indirect supervision applied to COVID-19 and pneumonia classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The COVID-19 pandemic necessitates advanced diagnostic tools.
• Machine learning offers data-driven solutions for healthcare challenges.
• Chest X-rays are crucial for diagnosing respiratory illnesses like COVID-19.

Purpose of the Study:

• To develop a Convolutional Neural Network (CNN)-based method for detecting COVID-19 using chest X-ray images.
• To address data scarcity by consolidating public datasets.
• To enhance classification accuracy through indirect supervision.

Main Methods:

• Utilized a CNN architecture incorporating convolutional units.
• Employed Grad-CAM for indirect supervision, using attention heatmaps to guide predictions.
• Aggregated and annotated data from five public sources into normal, pneumonia, and COVID-19 categories.
• Implemented a training pipeline based on indirect supervision of traditional classification networks.

Main Results:

• Achieved high classification accuracy for COVID-19 detection.
• Demonstrated that standard CNNs can perform comparably to tailor-made models.
• VGG-16, a widely used network, outperformed specialized models in COVID-19 detection.
• Grad-CAM-guided indirect supervision effectively supported network predictions.

Conclusions:

• A CNN-based approach with indirect supervision shows promise for COVID-19 detection from chest X-rays.
• Consolidating diverse datasets is effective in overcoming data scarcity.
• Standard, well-established networks can be highly effective for specific medical imaging tasks.
• Indirect supervision using techniques like Grad-CAM can enhance the performance of diagnostic AI models.