Rectifying Supporting Regions With Mixed and Active Supervision for Rib Fracture Recognition

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Automatic rib fracture recognition from chest X-rays is crucial but challenging due to subtle fracture indicators.
• Weakly Supervised Learning (WSL) models use image-level labels but Class Activation Maps (CAMs) may focus on irrelevant regions, impacting clinical reliability.
• Existing Mixed Supervised Learning (MSL) models require extensive object-level annotations, which are scarce for rib fractures.

Purpose of the Study:

• To propose a novel MSL framework for robust and reliable automatic rib fracture recognition in chest X-rays.
• To enhance the accuracy and interpretability of WSL-derived CAMs by guiding them towards true fracture locations.
• To minimize the annotation cost in MSL by incorporating an active learning strategy.

Main Methods:

• Developed a novel MSL framework integrating adversarial classification learning into WSL.
• Introduced Biased Correlation Decoupling and Instance Separation Enhancing strategies to indirectly guide CAMs to fractures.
• Implemented a CAM-based Active Learning strategy to identify and prioritize samples for annotation, reducing overall annotation burden.

Main Results:

• The proposed method generated rational Supporting Regions for CAMs, improving interpretability of classification decisions.
• Experimental results on a large chest X-ray dataset demonstrated superior performance compared to competing methods.
• Achieved high performance with only 20% of positive samples annotated with bounding boxes, significantly reducing annotation costs.

Conclusions:

• The novel MSL framework effectively addresses the challenges of rib fracture recognition in chest X-rays.
• The proposed strategies enhance the reliability and interpretability of WSL models for clinical applications.
• This approach offers a cost-effective solution for training accurate rib fracture detection models by minimizing annotation requirements.