Ensemble of LinkNet Networks for Head and Neck Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiotherapy Planning

Background:

• Accurate segmentation of head and neck cancer (HNC) tumors is crucial for effective radiotherapy treatment planning.
• Manual segmentation is time-consuming and prone to inter-observer variability.
• Automated segmentation algorithms offer a potential solution to streamline the radiation oncology process.

Purpose of the Study:

• To develop and evaluate an ensemble of LinkNet networks for automated head and neck cancer tumor segmentation.
• To participate in the HNTS-MRG 2024 Grand Challenge for HNC tumor segmentation.
• To assess the performance of ensemble learning in improving segmentation accuracy compared to individual models.

Main Methods:

• An ensemble of eight 2D LinkNet networks was created using selected weights from a single pretrained LinkNet model.
• The individual LinkNet networks were trained on the HNC dataset for 200 epochs.
• Predictions from the eight networks were averaged to generate the final segmentation mask.

Main Results:

• The LinkNet Ensemble achieved an aggregated Dice score of 64.60% for metastatic lymph nodes and 49.53% for primary gross tumors.
• The ensemble model outperformed individual LinkNet architectures.
• The overall mean score for the challenge was 57.06%.

Conclusions:

• Ensemble learning effectively enhances head and neck cancer tumor segmentation accuracy without significantly increasing computational costs.
• The developed automated segmentation method shows promise for improving radiotherapy treatment planning efficiency and precision.
• The LinkNet Ensemble demonstrated competitive performance in the HNTS-MRG 2024 Grand Challenge.