Enhancing Head and Neck Tumor Segmentation in MRI: The Impact of Image Preprocessing and Model Ensembling.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Online adaptive MR-guided radiotherapy (MRgRT) is crucial for Head and Neck Cancer (HNC) treatment.
• Accurate tumor delineation is a significant challenge in HNC radiotherapy planning.
• Manual segmentation is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and optimize automated segmentation models for HNC tumors in MR images.
• To evaluate the impact of preprocessing techniques and model ensembling on segmentation accuracy.
• To provide an effective solution for the MICCAI Head and Neck Tumor Segmentation for MR-Guided Applications (HNTS-MRG) challenge.

Main Methods:

• Investigated various preprocessing techniques including maximal cropping and contrast enhancement.
• Developed and ensembled robust deep learning segmentation models.
• Validated models on internal datasets and submitted to the HNTS-MRG challenge for pre-RT and mid-RT tasks.

Main Results:

• Ensembled models achieved high Dice scores, with internal validation yielding (0.680, 0.785) for GTVp/GTVn on pre-RT and (0.493, 0.810) on mid-RT.
• Submitted models under the team name "Stockholm_Trio" achieved aggregated Dice scores of (0.795, 0.849) for pre-RT and (0.553, 0.865) for mid-RT.
• The developed models demonstrated superior performance in HNC tumor segmentation.

Conclusions:

• Optimized preprocessing and model ensembling are effective for improving HNC tumor segmentation accuracy.
• The developed automated segmentation approach supports the adoption of online adaptive MRgRT.
• The study provides a valuable, publicly available resource for HNC tumor segmentation research.