Accurate Screening of COVID-19 Using Attention-Based Deep 3D Multiple Instance Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate COVID-19 screening from chest CT scans remains challenging due to 3D spatial complexity, labeling difficulties, and subtle differences from other pneumonias.
• Existing methods often require manual annotation of infection areas or lack interpretability.
• The COVID-19 pandemic highlighted the urgent need for efficient and reliable automated screening tools.

Purpose of the Study:

• To develop a highly accurate and interpretable method for screening COVID-19 from chest CT scans using weak labels.
• To address the limitations of existing automated screening approaches.
• To provide an efficient assisted tool for COVID-19 diagnosis.

Main Methods:

• Proposed an attention-based deep 3D multiple instance learning (AD3D-MIL) framework.
• Treated 3D chest CT scans as bags of instances, using patient-level labels.
• Implemented attention-based pooling for instance contribution analysis and learned Bernoulli distributions for bag-level labels.

Main Results:

• Achieved an overall accuracy of 97.9% on a dataset of 460 chest CT examples.
• Demonstrated high performance with an AUC of 99.0% and a Cohen kappa score of 95.7%.
• The AD3D-MIL model successfully generated interpretable 3D instances and provided insights into their contribution.

Conclusions:

• The developed AD3D-MIL algorithm offers a robust and interpretable solution for automated COVID-19 screening from chest CT scans.
• The model's high accuracy and interpretability make it a valuable assisted diagnostic tool.
• This approach effectively utilizes weak labels, overcoming a significant challenge in medical image analysis.