Multimodality Multimask and Multitask Auto-Segmentation Network for Organs-at-Risk in Head and Neck Radiation Therapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate segmentation of head and neck organs-at-risk is crucial for radiation therapy planning.
• Current single-modality segmentation methods struggle with soft-tissue differentiation and patient variations.

Purpose of the Study:

• To develop a clinically robust auto-segmentation framework integrating multimodal imaging.
• To optimize computational efficiency in head and neck cancer treatment planning.

Main Methods:

• Developed M3-Net, a deep learning architecture with cross-modality fusion and attention mechanisms.
• Employed a hierarchical multimask generator and dual-task learning (segmentation + deformable image registration).
• Trained on 200 retrospective and 10 prospective cases.

Main Results:

• M3-Net reduced inference time by 63.6% and achieved Dice similarity coefficient >0.88 for 60% of organs.
• The framework demonstrated superior performance in 93.3% of organs with the best average surface distance.
• Clinical validation confirmed speed and precision meet requirements for radiation therapy.

Conclusions:

• M3-Net sets a new standard for head and neck organs-at-risk segmentation, balancing accuracy and efficiency.
• The clinically validated workflow significantly reduces contouring time while maintaining precision.
• Enables faster adoption of adaptive radiation therapy protocols.