Deep Learning Methods for Chest Disease Detection Using Radiography Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• X-ray imaging is a primary diagnostic tool for various diseases.
• Deep learning (DL) computer-aided detection (CAD) systems enhance radiologist capabilities.
• Accurate classification of chest diseases from X-rays is crucial.

Purpose of the Study:

• To propose a novel two-step deep learning approach for chest disease classification.
• To evaluate two new DL models, DC-ChestNet and VT-ChestNet, on a large CXR dataset.
• To compare the proposed models against existing state-of-the-art methods.

Main Methods:

• A two-step classification strategy: multi-class (normal, lung, heart) followed by binary classification of specific diseases.
• Development of DC-ChestNet (ensemble of DCNNs) and VT-ChestNet (modified transformer).
• Training and validation on a consolidated dataset of 26,316 chest X-ray (CXR) images.

Main Results:

• VT-ChestNet demonstrated superior performance compared to DC-ChestNet and other leading models.
• VT-ChestNet achieved an Area Under Curve (AUC) of 95.13% in the initial multi-class classification step.
• For the second step, VT-ChestNet attained average AUCs of 99.26% for heart diseases and 99.57% for lung diseases.

Conclusions:

• The proposed two-step approach, particularly VT-ChestNet, shows significant potential for accurate chest disease classification from X-rays.
• Transformer-based models offer a promising direction for advancing medical image analysis in radiology.
• VT-ChestNet provides a robust and high-performing solution for computer-aided detection of lung and heart conditions.