Human-level COVID-19 diagnosis from low-dose CT scans using a two-stage time-distributed capsule network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• Reverse transcription-polymerase chain reaction (RT-PCR) is the standard for COVID-19 diagnosis but has limitations including delayed results and high false negative rates.
• Chest computed tomography (CT) aids in COVID-19 diagnosis and assessment, but standard doses pose a significant radiation risk, especially for patients requiring repeat scans.
• Low-dose and ultra-low-dose CT (LDCT/ULDCT) protocols reduce radiation exposure to near X-ray levels while preserving diagnostic resolution.

Purpose of the Study:

• To develop and evaluate an Artificial Intelligence (AI)-based framework for diagnosing COVID-19 using LDCT/ULDCT scans.
• To determine if the AI model can achieve human-level diagnostic performance, compensating for potential shortages in thoracic radiology expertise during the pandemic.
• To assess the impact of incorporating clinical data alongside imaging for improved diagnostic accuracy.

Main Methods:

• Development of a two-stage capsule network architecture for the AI model.
• Training and validation of the AI model on a dataset of LDCT/ULDCT scans to classify COVID-19, community-acquired pneumonia (CAP), and normal cases.
• Cross-validation to assess model performance metrics including sensitivity and accuracy, both with and without clinical data integration.

Main Results:

• The AI model achieved high sensitivity for COVID-19 ([Formula: see text]) and CAP ([Formula: see text]), with excellent specificity for normal cases ([Formula: see text]) and overall accuracy ([Formula: see text]) using LDCT/ULDCT scans.
• Incorporating clinical data (demographics, symptoms) further improved performance, yielding COVID-19 sensitivity of [Formula: see text], CAP sensitivity of [Formula: see text], normal case specificity of [Formula: see text], and accuracy of [Formula: see text].
• The AI model demonstrated human-level diagnostic capabilities on reduced radiation dose scans.

Conclusions:

• The proposed AI model offers a promising tool for rapid and accurate COVID-19 diagnosis using LDCT/ULDCT scans with significantly reduced radiation exposure.
• The AI framework has the potential to support radiologists, improve diagnostic turnaround times, and aid in controlling disease transmission during pandemics.
• This approach addresses the need for accessible diagnostic tools in situations with limited access to specialized radiological expertise.