Deep learning for segmentation in radiation therapy planning: a review.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Manual segmentation in radiation therapy is time-consuming and prone to inter-observer variability.
• Deep learning, especially convolutional neural networks (CNNs), presents advanced auto-segmentation solutions.
• Accurate segmentation of organs and structures is crucial for effective radiation therapy.

Purpose of the Study:

• To provide a descriptive review of deep learning techniques for segmentation in radiation therapy planning.
• To identify common architectures, data sets, and methodologies in the field.
• To highlight areas for improvement in research and clinical application.

Main Methods:

• Literature review of deep learning segmentation methods in radiation therapy.
• Analysis of common CNN architectures, focusing on U-net.
• Examination of data sets (CT images) and validation strategies (N-fold cross-validation).

Main Results:

• U-net emerged as the most prevalent CNN architecture.
• Head and neck normal tissue segmentation was the most common application.
• CT images were the primary data source, with mixed use of in-house and public datasets.
• N-fold cross-validation was frequently used, but data separation for training, testing, and validation was inconsistent.

Conclusions:

• Deep learning shows significant promise for automating segmentation in radiation therapy.
• Standardization of metrics and independent validation are essential for method comparison and benchmarking.
• The field is rapidly evolving, necessitating continued research and development.