Head and neck tumor segmentation convolutional neural network robust to missing PET/CT modalities using channel...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology

Background:

• Accurate gross tumor volume segmentation is crucial for effective head and neck (H&N) cancer radiation therapy.
• Current segmentation methods may lack robustness when dealing with single imaging modalities.

Purpose of the Study:

• To develop a novel deep learning segmentation model for H&N cancer.
• To leverage combined CT and FDG-PET imaging data for improved segmentation accuracy.

Main Methods:

• A 3D U-Net architecture with deep supervision was implemented.
• Channel dropout was used to ensure model robustness with single or combined imaging modalities.
• Ensemble modeling combined conventional and dilated convolutions for enhanced feature capture.

Main Results:

• The model achieved a Dice Similarity Coefficient (DSC) of 0.802 for combined CT and PET.
• Performance was also strong on single modalities: DSC of 0.610 for CT and 0.750 for PET.
• The channel dropout technique enabled high performance across all tested modality combinations.

Conclusions:

• The developed deep learning model offers a robust and accurate solution for H&N cancer gross tumor volume segmentation.
• The model's adaptability to single or combined imaging modalities enhances its clinical relevance.
• This approach supports therapeutic management in scenarios where specific imaging data may be unavailable.