Comparison of Human-in-the-Loop Neural Network and Manual Methods for Aortic Diameter Measurement at CT Angiography

Area of Science:

• Medical Imaging and Artificial Intelligence
• Cardiovascular Imaging Analysis
• Deep Learning in Radiology

Background:

• Accurate measurement of thoracic aortic diameter is crucial for managing aortic dilatation.
• Manual measurements can be time-consuming and prone to variability.
• Deep learning models offer potential for automated and precise image analysis.

Purpose of the Study:

• To compare a U-Net convolutional neural network with human-in-the-loop (HITL) validation against manual clinical measurements for thoracic aortic diameter.
• To evaluate the performance of automated segmentation and measurement of the thoracic aorta using CT angiography.

Main Methods:

• Retrospective analysis of CT angiographic examinations in patients with thoracic aortic dilatation.
• Automated aortic segmentation, landmark localization, and diameter measurements using a multitask U-Net.
• HITL validation of U-Net measurements, with expert remeasurement for discrepancies > 5 mm.
• Comparison of mid ascending aortic growth using U-Net, manual, and vascular deformation mapping methods.

Main Results:

• 96% of automated measurements passed HITL validation.
• HITL validation significantly reduced measurement discrepancies and improved limits of agreement.
• U-Net measurements showed stronger agreement with a diameter-independent reference standard (vascular deformation mapping) than manual measurements.

Conclusions:

• Automated U-Net measurements, enhanced by HITL validation, provide more accurate thoracic aortic diameter assessments.
• This AI-assisted approach demonstrates superior agreement with reference standards compared to traditional manual measurements.
• The findings support the clinical utility of AI for improved thoracic aorta evaluation in CT angiography.