CT-based COPD identification using multiple instance learning with two-stage attention | JoVE Visualize

Area of Science:

Medical Imaging and Diagnostics
Artificial Intelligence in Healthcare
Pulmonary Medicine

Background:

Chronic obstructive pulmonary disease (COPD) is a global health burden, often underdiagnosed due to limitations in current diagnostic tools like spirometry for mild cases.
Chest computed tomography (CT) offers a validated method for COPD diagnosis and quantification, but deep learning approaches face challenges in characterizing its complex, heterogeneous pathological alterations.
Improving the diagnostic performance of CT-based methods is crucial for effective COPD management.

Purpose of the Study:

To develop and validate a novel deep learning approach for accurate COPD identification using CT images.
To address the challenges of multidimensional and spatially heterogeneous pathological alternations in COPD.
To enhance the diagnostic performance beyond existing CT-based deep learning methods.

Main Methods:

A two-stage attention multiple instance learning (TSA-MIL) framework was developed, utilizing a Resnet-50 model pre-trained on natural images for feature extraction.
A pseudo-color method was employed to enhance feature representation from single-channel CT slices, and a two-stage attention module was designed to refine instance-level attention scores.
Instance-level clustering was used to improve feature separability, with the model trained and validated on data from 800 participants and further tested on an external dataset of 260 participants.

Main Results:

The TSA-MIL approach achieved high diagnostic accuracy, with an accuracy of 0.9200 and an AUC of 0.9544 on the test set.
The model demonstrated strong performance on an external validation set, achieving an accuracy of 0.8115 and an AUC of 0.8737.
The proposed method outperformed advanced MIL models and other state-of-the-art COPD identification techniques.

Conclusions:

The TSA-MIL approach represents a significant advancement in CT-based COPD identification, offering clinically acceptable diagnostic performance.
This method shows potential as a powerful tool to aid clinicians in the diagnosis of COPD.
Further validation and implementation could improve early detection and management of chronic obstructive pulmonary disease globally.