Strategies to improve deep learning-based salivary gland segmentation

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Manual delineation of organs-at-risk in radiotherapy is time-consuming and variable.
• Deep learning offers a potential solution to improve efficiency and consistency.
• Salivary glands serve as a model system for evaluating deep learning segmentation.

Purpose of the Study:

• To systematically evaluate strategies for enhancing deep learning performance and reliability in organ-at-risk segmentation.
• To identify optimal methods for improving salivary gland segmentation for radiotherapy applications.

Main Methods:

• Investigated effects of increasing training data, data augmentation (traditional and domain-specific), data quality, custom cost functions, patient-specific Hounsfield unit windowing, and model ensembles.
• Measured performance using geometric parameters and reliability using parameter variance.

Main Results:

• Increasing training data, data augmentation, patient-specific windowing, and model ensembles positively impacted performance and reliability.
• Combined strategies improved Sørensen-Dice coefficient by 3-4% and decreased standard deviation by 1% for salivary glands.
• Performance gains diminished with already high base model performance.

Conclusions:

• Specific strategies significantly enhance deep learning model performance and reliability for organ-at-risk segmentation.
• Clinical adoption of automated salivary gland segmentation is facilitated by reduced post-segmentation editing.
• Optimized deep learning models can improve radiotherapy planning and adaptive treatments.