Enhancing automatic landmark localization in X-ray images using combined segmentation and regression models:...

Area of Science:

• Medical Imaging
• Computer Vision
• Orthopedics

Background:

• Manual landmark detection in lower limb medical imaging is inefficient and prone to errors.
• Automatic methods combining image segmentation and coordinate regression show promise but require optimization.
• Design choices, integration strategies, and hyperparameter tuning are critical for performance.

Purpose of the Study:

• To investigate the optimal approach for combining image segmentation and coordinate regression for lower limb landmark detection.
• To compare network architectures and training strategies for improved accuracy and robustness.
• To evaluate the clinical utility of automated detection in lower limb malalignment assessment.

Main Methods:

• Compared U-Net and Swin-UNETR for landmark segmentation.
• Assessed VGG-16, ResNet-50, and Swin-B for coordinate regression.
• Evaluated end-to-end training versus cascading subnetworks.
• Tested performance on lower-limb X-rays and hip-knee-ankle angle measurements.

Main Results:

• Swin-UNETR slightly outperformed U-Net in segmentation accuracy (lower Euclidean distance error) and false positives.
• VGG-16 with end-to-end training achieved the highest coordinate regression accuracy.
• The combined Swin-UNETR and VGG-16 end-to-end framework demonstrated superior performance.

Conclusions:

• An end-to-end framework combining Swin-UNETR and VGG-16 is optimal for automated lower limb landmark detection.
• This approach significantly improves accuracy and robustness in clinical malalignment assessment.
• Automated methods offer a more efficient and reliable alternative to manual landmark detection.