Dual-level weighted cross-entropy loss function and multi-object region segmentation network evaluation for dynamic...

Area of Science:

• Medical imaging
• Computer vision
• Biomedical engineering

Background:

• The knee joint is crucial for body support and movement, necessitating accurate assessment via dynamic X-ray radiography.
• Automatic segmentation of knee structures (patella, femur, tibia, patellar tendon) in dynamic X-rays is vital for diagnostic efficiency.
• Network architecture and loss functions are key to developing effective segmentation models.

Purpose of the Study:

• To propose an optimal multi-object region segmentation model for dynamic knee joint X-ray radiography.
• To enhance the accuracy and efficiency of quantitative analysis of knee joint motion.

Main Methods:

• Developed a dual-level weighted cross-entropy loss function to balance segmentation losses across multiple knee structures.
• Created two comprehensive evaluation metrics for multi-object region segmentation models.
• Proposed a scoring criterion to determine the optimal segmentation model and loss function ratios.

Main Results:

• The proposed dual-level weighted cross-entropy loss function improved segmentation performance compared to traditional methods.
• The optimal model (DeepLabV3+_R50c with a specific mixed loss function) achieved high segmentation accuracy (mean IoU: 0.8921, mean Dice: 0.9373).
• Achieved excellent performance metrics including Precision (0.9316), Recall (0.9490), HD95 (2.9145), and ASSD (1.0309).

Conclusions:

• The developed multi-object region segmentation model significantly enhances the accuracy of knee joint motion analysis.
• This approach has the potential to improve diagnostic processes and reduce labor for radiologists and physicians.