COVID-19 detection in CT images with deep learning: A voting-based scheme and cross-datasets analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Control

Background:

• Early detection of COVID-19 is crucial for controlling its spread.
• Deep learning models are being developed for automated COVID-19 screening using CT scans.
• Existing methods often treat CT slices independently and are trained/tested on single datasets, limiting generalizability.

Purpose of the Study:

• To propose an efficient deep learning technique with a voting-based approach for COVID-19 screening in CT scans.
• To address the limitations of independent slice analysis and single-dataset training.
• To evaluate the model's performance and generalization capabilities across different datasets.

Main Methods:

• A voting-based deep learning approach was developed to classify patient CT scans as a group.
• The proposed method was tested on two large COVID-19 CT datasets with a patient-based split.
• Cross-dataset studies were conducted to assess model robustness and generalization in realistic scenarios.

Main Results:

• The voting-based approach improved patient-level classification.
• Cross-dataset analysis revealed a significant drop in accuracy from 87.68% to 56.16% in the best-case scenario.
• This indicates poor generalization power of current deep learning models for COVID-19 CT screening.

Conclusions:

• Current deep learning methods for COVID-19 detection in CT images require significant improvement for clinical application.
• The study highlights the inadequacy of existing models in handling data from different distributions.
• Larger and more diverse datasets are essential for robust evaluation and clinical adoption of AI-based diagnostic tools.