COVID-19 early detection for imbalanced or low number of data using a regularized cost-sensitive CapsNet

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic necessitated rapid diagnostic tools, with deep learning showing promise in analyzing lung CT scans.
• Current deep learning methods for COVID-19 detection often require large datasets, limiting their applicability in real-world clinical settings with scarce data.

Purpose of the Study:

• To develop a robust deep learning architecture for COVID-19 detection from CT images that addresses the challenge of limited and imbalanced data.
• To enhance the generalization and convergence of deep learning models for improved diagnostic performance.

Main Methods:

• A hybrid deep learning architecture combining DenseNet and CapsNet was proposed.
• A novel regularization term with fewer parameters was introduced to improve model generalization.
• A cost-sensitive loss function was developed to address imbalanced datasets.

Main Results:

• The proposed model demonstrated improved network convergence, particularly with small training datasets.
• The approach showed superior performance on imbalanced data compared to existing methods.
• The model achieved state-of-the-art results on the HUST and COVID-CT datasets, outperforming previous benchmarks.

Conclusions:

• The combined DenseNet-CapsNet architecture with proposed enhancements offers a potent solution for COVID-19 detection using CT images.
• The method is particularly effective in real-world hospital scenarios characterized by limited and imbalanced data.
• This approach advances computer-aided diagnosis for infectious respiratory diseases.