Deep Learning-Based Object Detection Strategies for Disease Detection and Localization in Chest X-Ray Images.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Chest X-ray (CXR) interpretation is crucial for diagnosing respiratory and cardiovascular diseases.
• Manual CXR analysis faces challenges including subjectivity, time consumption, and diagnostic errors.
• Automated methods are needed to improve accuracy and consistency in CXR interpretation.

Purpose of the Study:

• To develop and evaluate deep learning-based object detection methods for automated identification and annotation of abnormalities in CXR images.
• To assess the impact of background image proportions on model performance.
• To compare Convolutional Neural Networks (CNNs) and Transformer-based models for medical image analysis.

Main Methods:

• Utilized disease-labeled CXR images with bounding boxes from E-Da Hospital for model development and testing.
• Investigated various training datasets and approaches to manage the prevalence of normal images.
• Explored few-shot object detection techniques to address limited data for specific diseases.
• Compared the performance of CNN and Transformer architectures.

Main Results:

• The proportion of background images significantly influenced model inference.
• Incorporating binary classification schemes consistently enhanced model performance.
• CNN-based models demonstrated superior performance compared to Transformer-based models across all tested scenarios.

Conclusions:

• A more efficient and reliable system for automated detection of disease labels and bounding boxes in CXR images has been developed.
• Deep learning, particularly CNNs, offers a promising approach to enhance CXR analysis.
• The study provides insights into optimizing deep learning models for medical image analysis, considering data imbalance and architectural choices.