Parotid Gland Segmentation Using Purely Transformer-Based U-Shaped Network and Multimodal MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Parotid gland tumors represent 2-10% of head and neck tumors, necessitating accurate segmentation on MRI for diagnosis and surgical planning.
• Parotid gland segmentation is challenging due to variable shapes and low contrast with surrounding tissues.
• Transformer networks show promise in computer vision but are underutilized for parotid gland segmentation.

Purpose of the Study:

• To develop and evaluate a purely Transformer-based U-shaped network for parotid gland and tumor segmentation on multi-center, multimodal MRI data.
• To improve segmentation accuracy and efficiency while reducing the manual labeling burden for clinicians.

Main Methods:

• Collected a multi-center, multimodal parotid gland MRI dataset.
• Implemented a Transformer-based U-shaped segmentation network incorporating absolute and relative positional encoding for enhanced feature learning.
• Developed a novel training approach to minimize clinician labeling effort.

Main Results:

• Achieved high segmentation performance with a Dice-Similarity Coefficient of 86.99%, Pixel Accuracy of 99.19%, Mean Intersection over Union of 81.79%, and Hausdorff Distance of 3.87.
• Demonstrated superior performance compared to convolutional neural networks.
• Successfully fused multimodal information from various MRI centers without increasing computational load.

Conclusions:

• The proposed Transformer-based U-shaped network offers a robust and accurate solution for parotid gland and tumor segmentation.
• The method effectively handles multimodal MRI data and significantly reduces the need for extensive manual annotation.
• This approach holds potential for improving diagnostic accuracy and treatment planning for parotid gland tumors.