Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy: Deep Learning Algorithm Development and...

Area of Science:

• Medical imaging and radiation oncology
• Artificial intelligence in healthcare
• Computational anatomy

Background:

• Head and neck cancer affects over 500,000 individuals globally each year.
• Radiotherapy is a key treatment, but manual organ delineation is time-consuming and variable.
• Automated segmentation tools face challenges in achieving expert-level performance.

Purpose of the Study:

• To develop and validate a 3D U-Net deep learning model for expert-level auto-segmentation of 21 head and neck organs at risk.
• To assess the clinical applicability and generalizability of the deep learning model.

Main Methods:

• A 3D U-Net architecture was trained on 663 computed tomography (CT) scans with expert segmentations.
• The model's performance was evaluated on a test set of 21 CT scans with dual expert delineations.
• A novel surface Dice similarity coefficient metric was introduced to assess contour accuracy.

Main Results:

• The deep learning model achieved expert-level performance in delineating 21 head and neck organs at risk.
• The surface Dice similarity coefficient quantified contour deviation, reflecting clinical error correction needs.
• The model demonstrated generalizability across diverse, open-source datasets.

Conclusions:

• Deep learning offers an effective and clinically applicable solution for head and neck auto-segmentation in radiotherapy.
• This technology has the potential to enhance the efficiency, consistency, and safety of radiotherapy workflows.
• Further validation and regulatory approval are needed for widespread clinical adoption.