CACTUSS: Common Anatomical CT-US Space for US examinations

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Vascular Surgery

Background:

• Abdominal aortic aneurysm (AAA) detection is challenging due to asymptomatic development and limitations of ultrasound (US) imaging.
• Computed tomography (CT) offers superior image quality for AAA monitoring but is less accessible for initial screening.
• Deep neural networks trained on CT data show promise for aorta segmentation.

Purpose of the Study:

• To leverage CT-derived labels to enhance aorta segmentation in ultrasound (US) images.
• To reduce the reliance on manual annotations for training US-based segmentation models.
• To improve the accuracy and efficiency of AAA screening and diagnosis using US.

Main Methods:

• Introduction of CACTUSS, a common anatomical CT-US space creating an intermediate representation (IR) space.
• Utilizing a re-parametrized physics-based US simulator to generate IR images for training.
• Employing an image-to-image translation network for application on real B-mode US images.

Main Results:

• CACTUSS-trained models demonstrated superior aorta segmentation performance compared to fully supervised methods.
• Quantitative evaluation using Dice Score and diagnostic metrics confirmed the model's effectiveness.
• The segmentation accuracy achieved meets clinical requirements for AAA screening and diagnosis.

Conclusions:

• CACTUSS offers a promising strategy for enhancing US segmentation accuracy by utilizing CT labels.
• The generated IR images preserve anatomical structures and are optimized for aorta segmentation.
• Future research includes integration into robotic US platforms for automated screening and clinical validation.