CXR-RefineDet: Single-Shot Refinement Neural Network for Chest X-Ray Radiograph Based on Multiple Lesions Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Radiologist workload increased during the COVID-19 pandemic, leading to diagnostic errors.
• Artificial intelligence (AI) offers potential for lesion detection in medical images.
• Current object detection models may not meet the demands for accurate and efficient lung disease diagnosis.

Purpose of the Study:

• To develop an advanced lung disease detection neural network for improved accuracy in medical imaging.
• To create a superior object detection model that addresses limitations of current mainstream approaches.

Main Methods:

• Designed a novel backbone network, RRNet, integrating RepVGG and Resblock for efficient feature extraction and fusion.
• Introduced an 'Information Reuse' structure to enhance feature utilization by feeding normalized features back into the network.
• Developed the CXR-RefineDet network by combining RRNet with an improved RefineDet architecture.

Main Results:

• CXR-RefineDet achieved a mean Average Precision (mAP) of 0.1686 on the VinDr-CXR dataset.
• The model demonstrated a fast inference speed of 6.8 frames per second (fps).
• Performance surpassed two-stage object detection algorithms utilizing strong backbones like ResNet-50 and ResNet-101.

Conclusions:

• CXR-RefineDet offers superior detection accuracy and inference speed compared to existing methods for lung disease detection.
• The model's efficiency makes it suitable for practical implementation in computer-aided diagnosis systems.
• AI-powered solutions like CXR-RefineDet can significantly support radiologists in diagnosing lung conditions.