Transformer-based and CNN-based models for clinically effective 2D and 3D pelvic bone segmentation in CT imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Accurate segmentation of pelvic structures in trauma CT scans is challenging due to anatomical complexity and image variability.
• Automated segmentation methods are crucial for efficient and precise analysis of pelvic fractures in clinical settings.

Purpose of the Study:

• To systematically compare Convolutional Neural Network (CNN) and transformer-based architectures for automated pelvic bone segmentation.
• To evaluate 2D and 3D model performance using different encoder backbones on the PENGWIN MICCAI 2024 dataset.
• To identify optimal deep learning strategies for clinical trauma imaging of pelvic fractures.

Main Methods:

• Utilized CT data from 150 patients with pelvic fractures from the PENGWIN MICCAI 2024 challenge dataset.
• Implemented and compared U-Net, LinkNet (CNNs), and UNETR (transformer) architectures in 2D and 3D formats with VGG19 and ResNet50 backbones.
• Trained models using composite loss functions and evaluated performance with 5-fold cross-validation using Dice coefficient, IoU, accuracy, sensitivity, and specificity.

Main Results:

• U-Net with ResNet50 achieved the highest 2D segmentation performance (Dice 0.991, IoU 0.982).
• 3D U-Net with VGG19 demonstrated strong volumetric segmentation (Dice 0.9112).
• UNETR showed superior specificity (0.993) and fast inference (<1 min/case) but lower sensitivity (0.730) for complex fragments.

Conclusions:

• Deep learning models effectively segment pelvic fractures with high accuracy and computational efficiency.
• The presented models provide a strong foundation for clinical integration into automated trauma imaging systems.
• Further external validation and workflow assessment are recommended for clinical implementation.