A Light Deep Learning Algorithm for CT Diagnosis of COVID-19 Pneumonia

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Numerous artificial intelligence (AI) algorithms exist for detecting COVID-19 pneumonia in chest CT scans.
• Limited availability of source code hinders the assessment of out-of-distribution generalization in existing AI models.
• There is a need for accessible, efficient AI tools for COVID-19 diagnosis from CT imaging.

Purpose of the Study:

• To introduce Cimatec-CovNet-19, a novel, lightweight 3D convolutional neural network for COVID-19 detection in chest CT scans.
• To make the source code of the developed AI model publicly available to the scientific community.
• To evaluate the performance of Cimatec-CovNet-19 in identifying COVID-19 positive cases.

Main Methods:

• Developed Cimatec-CovNet-19, a 3D convolutional neural network inspired by VGG16 architecture.
• Trained the model on 3000 CT scans (1500 COVID-19 positive) globally, augmented with 3000 additional images.
• Implemented a novel pre-processing technique using lung CT masks for slice selection, requiring only 16 slices per examination.

Main Results:

• Achieved a recall of 0.88, specificity of 0.88, ROC-AUC of 0.95, PR-AUC of 0.95, and F1-score of 0.88 on a test set of 414 samples.
• Demonstrated effective COVID-19 identification using only 16 selected CT slices.
• The model proved to be a light and effective screening tool for COVID-19.

Conclusions:

• Cimatec-CovNet-19 is a promising, efficient, and accessible AI tool for screening COVID-19 from chest CT scans.
• The open-source nature of the code facilitates further research into AI generalization capabilities.
• The model's performance supports its utility in clinical settings for rapid COVID-19 assessment.