Weakly Supervised Object Detection in Chest X-Rays With Differentiable ROI Proposal Networks and Soft ROI Pooling

Area of Science:

• Medical imaging analysis
• Computer-aided diagnosis
• Machine learning for healthcare

Background:

• Weakly supervised object detection (WSup-OD) enhances image classification interpretability without extra annotations.
• Existing WSup-OD methods struggle with medical images due to unique object characteristics.

Purpose of the Study:

• To introduce a novel WSup-OD method for medical imaging.
• To improve disease localization in chest X-rays.

Main Methods:

• Proposed Weakly Supervised ROI Proposal Networks (WSRPN) with a region of interest-attention (ROI-attention) module.
• Integrated WSRPN with standard classification algorithms.
• End-to-end training using only image-level labels.

Main Results:

• WSRPN effectively generates bounding box proposals for pathologies.
• The method demonstrated superior performance compared to existing approaches.
• Achieved high accuracy in disease localization on chest X-ray datasets.

Conclusions:

• WSRPN offers a robust solution for weakly supervised disease localization in medical images.
• The proposed ROI-attention mechanism is key to its success.
• This method advances the application of WSup-OD in clinical settings.