CECT: Controllable ensemble CNN and transformer for COVID-19 image classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning

Background:

• The COVID-19 pandemic necessitates rapid and accurate diagnostic tools.
• Existing diagnostic methods may have limitations in speed and accuracy.

Purpose of the Study:

• To develop a novel deep learning network, CECT, for timely and accurate COVID-19 diagnosis.
• To enhance feature extraction capabilities for improved diagnostic performance.

Main Methods:

• Developed a Controllable Ensemble Convolutional neural network and Transformer (CECT) model.
• CECT utilizes parallel convolutional encoder, aggregate transposed-convolutional decoder, and windowed attention blocks.
• The model captures features at multi-local and global scales with controllable feature contribution.

Main Results:

• CECT achieved 98.1% accuracy in intra-dataset evaluation on public COVID-19 datasets.
• Demonstrated strong generalization with 90.9% accuracy on an unseen dataset.
• Outperformed existing state-of-the-art methods in COVID-19 classification.

Conclusions:

• CECT provides a powerful and accurate tool for COVID-19 diagnosis.
• The model's robust feature capture and generalization abilities suggest potential for broader medical applications.
• The developed CECT network can be extended to other medical diagnostic scenarios.