MRI-Based Head and Neck Tumor Segmentation Using nnU-Net with 15-Fold Cross-Validation Ensemble.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• MRI offers superior soft tissue contrast for tumor segmentation compared to CT and PET.
• Accurate tumor segmentation is vital for effective adaptive radiotherapy planning.
• The Head and Neck Tumor Segmentation for MR-Guided Applications (HNTSMRG-24) challenge addresses this need.

Purpose of the Study:

• To evaluate the performance of the nnU-Net V2 framework for segmenting primary gross tumor volume (GTVp) and metastatic lymph nodes (GTVn) in head and neck cancer.
• To compare segmentation accuracy at pre-radiotherapy (pre-RT) and mid-radiotherapy (mid-RT) stages.
• To enhance robustness through a 15-fold cross-validation ensemble.

Main Methods:

• Utilized the nnU-Net V2 framework with a 15-fold cross-validation ensemble.
• Augmented pre-RT segmentation data with corresponding mid-RT volumes.
• Employed a three-channel input for mid-RT segmentation, including registered pre-RT MRI and mask.

Main Results:

• Achieved an aggregated Dice Similarity Coefficient (DSC) of 0.81 for Task 1 (pre-RT segmentation) and 0.70 for Task 2 (mid-RT segmentation).
• Specific DSCs for Task 1 were 0.77 (GTVp) and 0.85 (GTVn).
• Specific DSCs for Task 2 were 0.54 (GTVp) and 0.86 (GTVn).

Conclusions:

• The proposed nnU-Net V2 approach demonstrates strong performance in head and neck tumor and lymph node segmentation.
• The method shows potential for improving adaptive radiotherapy treatment planning through accurate MR-based segmentation.
• The use of cross-validation ensembles and multi-channel inputs enhances segmentation accuracy and robustness.