A five-layer deep convolutional neural network with stochastic pooling for chest CT-based COVID-19 diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• The COVID-19 pandemic caused millions of cases globally by August 2020.
• Chest CT scans are a crucial diagnostic tool for identifying COVID-19.
• Existing diagnostic methods require improvement in accuracy and speed.

Purpose of the Study:

• To develop a novel deep learning model for accurate and swift COVID-19 diagnosis using chest CT images.
• To enhance a traditional deep convolutional neural network (DCNN) model with specific improvements.

Main Methods:

• Introduced stochastic pooling to replace traditional pooling methods (average and max pooling).
• Developed a 'conv block' (CB) by combining convolutional and batch normalization layers.
• Created a 'fully connected block' (FCB) by integrating dropout and fully connected layers.
• Configured the model as 3CB + 2FCB for optimal performance.

Main Results:

• The proposed model achieved high diagnostic performance: 93.28% sensitivity, 94.00% specificity, and 93.64% accuracy.
• Stochastic pooling demonstrated superior performance compared to average and max pooling.
• The 3CB + 2FCB configuration proved to be the most effective structure.

Conclusions:

• The novel deep learning model is effective for detecting COVID-19 from chest CT images.
• The proposed architectural improvements enhance diagnostic capabilities.
• This AI-driven approach offers a promising solution for rapid COVID-19 screening.