Head and Neck Tumor Segmentation from [18F]F-FDG PET/CT Images Based on 3D Diffusion Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Head and neck (H&N) cancers are a significant global health concern.
• Positron Emission Tomography/Computed Tomography (PET/CT) is crucial for H&N cancer management.
• Diffusion models show promise in image generation tasks.

Purpose of the Study:

• To develop and evaluate a 3D diffusion model for precise H&N tumor segmentation using PET and CT data.
• To compare the 3D diffusion model's performance against state-of-the-art methods.

Main Methods:

• A 3D diffusion model incorporating a 3D UNet architecture was developed.
• The model utilized concatenated PET, CT, and Gaussian noise volumes for tumor mask generation.
• Experiments were conducted on the HECKTOR challenge dataset, comparing against U-Net and Transformer-based methods.

Main Results:

• The proposed 3D diffusion model achieved more accurate H&N tumor segmentation compared to other methods.
• The 3D model demonstrated superior performance over its 2D counterpart.
• Utilizing dual-modality PET and CT data proved advantageous over single-modality data.

Conclusions:

• The 3D diffusion model offers a highly effective approach for H&N tumor segmentation.
• The study underscores the benefits of 3D modeling and multi-modal data in medical image analysis for oncology.