Advances in Auto-Segmentation.

Area of Science:

• Medical imaging
• Radiotherapy
• Computational anatomy

Background:

• Manual image segmentation in radiotherapy is crucial but time-consuming and prone to inter- and intraobserver variability.
• Accurate segmentation is essential for radiotherapy plan optimization, quality assessment, and subsequent analyses like radiomics and dosimetrics.
• Current auto-segmentation methods, including multiatlas and hybrid techniques, are considered state-of-the-art but may be surpassed by newer approaches.

Purpose of the Study:

• To review traditional (non-deep learning) auto-segmentation algorithms used in radiotherapy.
• To introduce deep learning concepts, specifically convolutional neural networks and fully-convolutional networks, for medical image segmentation.
• To summarize deep learning-based auto-segmentation applications in radiotherapy and discuss clinical deployment considerations.

Main Methods:

• Review of traditional image segmentation algorithms relevant to radiotherapy.
• Introduction to deep learning principles and architectures (CNNs, FCNs) for segmentation.
• Literature survey of deep learning auto-segmentation applications in radiotherapy.

Main Results:

• Traditional auto-segmentation methods have limitations in accuracy and consistency.
• Deep learning, particularly CNNs and FCNs, shows significant promise for advancing auto-segmentation in radiotherapy.
• A growing body of literature demonstrates the effectiveness of deep learning for segmenting targets and normal tissues in radiotherapy.

Conclusions:

• Deep learning represents the fourth generation of auto-segmentation algorithms, offering potential to overcome limitations of previous methods.
• Auto-segmentation using deep learning can improve the accuracy, efficiency, and consistency of radiotherapy planning.
• Considerations for clinical implementation, including commissioning and quality assurance, are vital for the successful deployment of auto-segmentation software.