ACSN: Attention capsule sampling network for diagnosing COVID-19 based on chest CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Automated computed tomography (CT) scan analysis aids rapid COVID-19 diagnosis, crucial for treatment and outbreak control.
• Existing capsule networks show promise but struggle with feature extraction from scattered lesions in CT scans.
• Current methods face challenges in effectively fusing multi-regional features due to sampling limitations.

Purpose of the Study:

• To propose an Attention Capsule Sampling Network (ACSN) for improved COVID-19 detection from chest CT scans.
• To address limitations in key slice extraction and feature fusion in existing diagnostic models.
• To enhance the accuracy and efficiency of automated COVID-19 diagnosis using CT imaging.

Main Methods:

• Developed an Attention Capsule Sampling Network (ACSN) integrating attention enhancement and dual sampling strategies.
• Implemented a key slices enhancement method to focus on critical information within numerous CT slices.
• Utilized two sampling types to retain both active and background features for comprehensive analysis.

Main Results:

• The proposed ACSN achieved high performance on an open dataset of 35,000 slices.
• Demonstrated superior results compared to state-of-the-art models in COVID-19 detection.
• Reported an accuracy of 96.3%, sensitivity of 98.8%, specificity of 93.8%, and AUC of 98.3%.

Conclusions:

• The ACSN model offers a significant advancement in automated COVID-19 detection using chest CT scans.
• The attention enhancement and integrated sampling methods effectively improve diagnostic accuracy.
• This approach holds potential for rapid and precise identification of COVID-19 cases, aiding clinical decision-making.