A Combined Loss-driven Framework for Automated Parotid Segmentation in Head-and-Neck Computed Tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Accurate parotid gland segmentation is crucial for radiotherapy planning to minimize radiation toxicity.
• Existing segmentation methods often struggle with anatomical accuracy and class imbalance.
• Deep learning offers a promising avenue for automated and precise medical image segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automatic parotid gland segmentation.
• To investigate the efficacy of 3D U-Net and attention-augmented 3D U-Net architectures.
• To assess a novel combined loss function integrating modified Dice Score Coefficient (mDSC) and focal loss (FL) for improved segmentation.

Main Methods:

• Utilized a dataset of 379 noncontrast head-and-neck CT scans with expert-verified contours.
• Implemented residual 3D U-Net and attention-enhanced 3D U-Net models using TensorFlow.
• Trained networks with a combined loss function (0.7 mDSC, 0.3 FL) and evaluated using DSC, IoU, and accuracy.

Main Results:

• The 3D U-Net with combined loss achieved a median Dice score of 0.8835 (left) and 0.8709 (right).
• Mean IoU values were 0.7672 (left) and 0.7358 (right), indicating strong segmentation performance.
• Bland-Altman analysis confirmed reduced variability and improved agreement, demonstrating robustness.

Conclusions:

• The combined loss function significantly enhances 3D U-Net segmentation performance, robustness, and spatial precision.
• The proposed framework demonstrates clinical feasibility for automated and reproducible parotid delineation.
• This deep learning approach holds potential for optimizing radiotherapy planning and reducing side effects.