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 common globally.
• [18F]F-FDG PET/CT is crucial for H&N cancer management.
• Diffusion models show promise in image-generation tasks.

Purpose of the Study:

• To propose and evaluate a 3D diffusion model for H&N tumor segmentation.
• To accurately segment tumors from 3D PET and CT volumes.
• To compare the 3D diffusion model against state-of-the-art methods.

Main Methods:

• Developed a 3D diffusion model incorporating a 3D UNet architecture.
• Utilized concatenated 3D PET, CT, and Gaussian noise volumes as input.
• Evaluated the model on the HECKTOR challenge dataset, comparing with U-Net and Transformer models.

Main Results:

• The 3D diffusion model achieved a mean Dice score of 0.739, outperforming other methods (Dice < 0.726).
• The 3D model significantly improved segmentation compared to a 2D diffusion model (Dice 0.739 vs. 0.669).
• Dual-modality PET/CT data yielded superior results (Dice > 0.570) over single-modality data.

Conclusions:

• The proposed 3D diffusion model demonstrates high effectiveness for H&N tumor segmentation.
• The 3D approach and dual-modality data integration enhance segmentation accuracy.
• This method offers a promising advancement for H&N cancer management.