Improving COVID-19 CT classification of CNNs by learning parameter-efficient representation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• The COVID-19 pandemic necessitates rapid and accurate diagnostic tools.
• Computed tomography (CT) imaging is vital for COVID-19 diagnosis.
• Existing deep learning models face challenges with data diversity, accuracy, and sensitivity.

Purpose of the Study:

• To enhance data diversity in COVID-19 CT datasets.
• To improve the accuracy and sensitivity of deep learning models for COVID-19 detection.
• To develop parameter-efficient representations for Convolutional Neural Networks (CNNs).

Main Methods:

• Designed incremental data augmentation techniques for CT images.
• Applied similarity regularization (SR) derived from contrastive learning.
• Trained and evaluated seven common CNNs on the COVIDx CT-2A dataset.

Main Results:

• Augmentation and SR techniques consistently improved CNN performance.
• DenseNet121 with SR achieved 99.44% accuracy in three-category classification.
• Achieved 98.40% precision, 99.59% sensitivity, and 99.50% specificity for COVID-19 pneumonia.

Conclusions:

• The proposed augmentation and SR methods significantly enhance COVID-19 detection accuracy.
• The approach surpasses existing state-of-the-art methods on the COVIDx CT-2A benchmark dataset.
• The developed techniques offer a promising direction for improving AI-assisted medical diagnosis.