FCF: Feature complement fusion network for detecting COVID-19 through CT scan images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate and timely diagnosis of COVID-19 is critical for effective quarantine and treatment.
• While RT-PCR is standard, computed tomography (CT) scans offer faster results, especially with AI assistance.
• Deep learning models can enhance COVID-19 detection from CT images, supporting clinical decision-making.

Purpose of the Study:

• To develop a deep learning classification model for detecting COVID-19 using lung CT images.
• To improve diagnostic accuracy and speed through an advanced AI framework.
• To create a model that complements existing diagnostic methods.

Main Methods:

• Proposed a Feature Complement Fusion network (FCF) utilizing both Convolutional Neural Networks (CNN) and Vision Transformer (ViT) extractors.
• Implemented a feature complement Transformer (FCT) with an attention mechanism to fuse local and global features effectively.
• Employed a combined supervised and weakly supervised training strategy to accelerate model convergence.

Main Results:

• Achieved a high accuracy of 99.34% on the test dataset.
• The FCF model demonstrated superior performance compared to current state-of-the-art classification models.
• The fusion of local and global features via FCT enhanced feature representation.

Conclusions:

• The FCF network provides a highly accurate and efficient method for COVID-19 detection from CT scans.
• The model's architecture effectively addresses feature extraction limitations by integrating CNN and ViT.
• The proposed framework shows potential for extension to other medical image classification tasks.