Self-Supervised Pre-Training with Intensity Guided Masking for Enhanced Aorta Segmentation in CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Vascular Surgery

Background:

• Abdominal aortic aneurysm (AAA) diagnosis and management rely on accurate aortic segmentation from CT imaging.
• Manual segmentation is labor-intensive, variable, and hinders clinical workflow.
• Existing deep learning methods require large annotated datasets, limiting their widespread use.

Purpose of the Study:

• To develop a self-supervised deep learning method for accurate aortic segmentation in CT scans.
• To reduce the dependency on large manually annotated datasets for training segmentation models.
• To improve the efficiency and reliability of assessing vascular conditions like AAA.

Main Methods:

• Proposed Intensity Guided Masking (IGM) for self-supervised pre-training of a deep learning model using CT image intensity properties.
• Integrated the pre-trained encoder into a SwinUNETR model for fine-tuning and aortic structure segmentation.
• Evaluated the method on public and private CT datasets.

Main Results:

• Achieved high segmentation accuracy with Dice Similarity Coefficient (DSC) of 91.20% on a public dataset and 85% on a private dataset.
• Reported low average surface distance (ASSD) of 0.05mm and 0.04mm on the respective datasets.
• Outperformed state-of-the-art supervised baselines and other pre-training techniques.

Conclusions:

• The IGM method effectively pre-trains deep learning models for aortic segmentation using self-supervision, reducing annotation burden.
• This approach enhances segmentation accuracy and efficiency for vascular condition assessment, particularly for AAA.
• The method shows significant potential for clinical application in improving patient care for AAA.