Multiview Self-Supervised Segmentation for OARs Delineation in Radiotherapy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate delineation of organs at risk (OARs) is crucial for effective head and neck (H&N) cancer radiotherapy.
• Manual OAR delineation is time-consuming and prone to inaccuracies.
• Deep learning models offer automated solutions but typically require extensive labeled data.

Purpose of the Study:

• To develop and evaluate a novel multiview contrastive representation learning method.
• To leverage unlabeled data to improve the performance of OAR delineation models.
• To address the challenge of limited labeled data in medical image analysis.

Main Methods:

• A multiview contrastive learning architecture was proposed, utilizing coronal, sagittal, and transverse CT views.
• A convolutional neural network processed the three projected 2D views to generate representations.
• Contrastive loss was applied to align representations from different views of the same image (positive pairs) and separate those from different images (negative pairs).

Main Results:

• The proposed method significantly improved quantitative performance in OAR delineation.
• Absolute Dice scores increased from 83% to 86% compared to state-of-the-art methods.
• The approach demonstrated effectiveness in a dataset of 220 H&N cancer patient CT images.

Conclusions:

• Multiview contrastive representation learning offers a powerful approach for enhancing OAR delineation models.
• This method effectively addresses the label-scarce problem in medical image segmentation.
• The findings suggest a principled way to utilize unlabeled data for improved radiotherapy planning.