Patch-based feature mapping with generative adversarial networks for auxiliary hip fracture detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Hip fractures are a major public health concern, especially in the elderly.
• Pelvic radiographs (PXRs) are essential for diagnosing hip fractures.
• Existing classification models for hip fracture detection sometimes lack explainability by focusing on non-fracture areas.

Purpose of the Study:

• To improve the explainability of hip fracture detection models using weakly supervised learning.
• To enhance the model's focus on the actual fracture region.
• To introduce a quantitative method for evaluating the model's focus on the region of interest (ROI).

Main Methods:

• Proposed a patch-auxiliary generative adversarial network (PAGAN) as an auxiliary module for classification models.
• Integrated PAGAN with SOTA models like EfficientNetB0, ResNet50, and DenseNet121.
• Utilized GradCAM for attention heatmaps and computed Intersection over Union (IOU) and Dice coefficient (Dise) to assess model explainability.

Main Results:

• PAGAN integration improved classification accuracy for EfficientNetB0 (93.61% to 95.97%), ResNet50 (90.66% to 94.89%), and DenseNet121 (93.51% to 94.49%).
• Model explainability, measured by IOU, significantly improved: EfficientNetB0 (0.32 to 0.54), ResNet50 (0.28 to 0.40), and DenseNet121 (0.37 to 0.51).

Conclusions:

• The proposed PAGAN module enhances both the performance and explainability of hip fracture detection models.
• PAGAN effectively directs model attention to the fracture region, improving diagnostic reliability.
• This approach offers a valuable tool for improving AI-driven medical image analysis in orthopedics.