A Graph Convolutional Multiple Instance Learning on a Hypersphere Manifold Approach for Diagnosing Chronic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Chronic obstructive pulmonary disease (COPD) is a leading cause of morbidity and mortality.
• Early diagnosis of COPD is crucial for effective treatment and improved quality of life.
• COPD manifestations in CT images are diverse, posing challenges for automated detection.

Purpose of the Study:

• To develop and validate a novel multiple instance learning (MIL) approach for accurate COPD diagnosis from CT images.
• To enhance the discriminative power of instance-level features for early-stage COPD detection.
• To leverage COPD severity information for improved diagnostic performance.

Main Methods:

• Proposed a graph convolutional MIL with adaptive additive margin loss (GCMIL-AAMS).
• Employed a self-attention mechanism within graph convolution and pooling for feature learning.
• Utilized a hypersphere manifold and adaptive angular margins for incorporating COPD severity grades.

Main Results:

• The GCMIL-AAMS method achieved high accuracy in COPD discrimination.
• Achieved AUCs of 0.960 ± 0.014 on the test set and 0.862 ± 0.010 on the external testing set.
• Demonstrated the applicability of graph learning to MIL for medical image analysis.

Conclusions:

• The GCMIL-AAMS approach offers superior discrimination and generalization for COPD diagnosis via CT.
• The study highlights the potential of integrating graph learning with MIL for complex medical conditions.
• This method can aid in the early and accurate detection of COPD, improving patient management.