Deep learning approach for the segmentation of aneurysmal ascending aorta

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Surgery

Background:

• Ascending thoracic aortic aneurysm (ATAA) diagnosis relies on maximum aortic diameter, which poorly predicts adverse events.
• Novel image-derived strategies are needed for individualized patient risk management.
• Deep learning shows promise for enhancing cardiovascular imaging analysis.

Purpose of the Study:

• To investigate the feasibility and efficacy of deep learning models (UNet, ENet, ERFNet) for automatic ATAA segmentation.
• To compare the accuracy and speed of different deep learning models in segmenting ATAAs.
• To assess the potential of deep learning in facilitating personalized ATAA management.

Main Methods:

• CT angiography data from 72 ATAA patients (tricuspid and bicuspid aortic valves) were semi-automatically segmented.
• UNet, ENet, and ERFNet deep learning models were trained using the segmented data.
• Segmentation performance was evaluated based on accuracy (Dice score) and inference time.

Main Results:

• All deep learning models achieved a Dice score >88%, indicating strong agreement with manual segmentation.
• ENet and UNet demonstrated higher accuracy compared to ERFNet.
• ENet exhibited significantly faster inference times than UNet.

Conclusions:

• Deep learning models can accurately and rapidly segment and quantify the 3D geometry of ATAAs.
• The findings support the integration of deep learning into clinical workflows for personalized ATAA management.
• This technology can improve risk stratification and patient care for ATAA.