Segment aorta and localize landmarks simultaneously on noncontrast CT using a multitask learning framework for...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Non-contrast chest CT scans, used for lung cancer screening, contain valuable data on the thoracic aorta.
• Assessing thoracic aorta morphology can aid in early detection of aortic diseases and predicting adverse events.
• Visual assessment of aortic morphology on these scans is challenging due to low contrast and relies heavily on physician expertise.

Purpose of the Study:

• To develop a novel deep learning framework for simultaneous segmentation and landmark localization of the thoracic aorta on unenhanced chest CT.
• To enable quantitative measurement of thoracic aorta morphology using the developed algorithm.

Main Methods:

• A multi-task deep learning framework with two subnets: one for aortic segmentation (sinuses, trunk, branches) and another for landmark detection (five key points).
• Shared encoder with parallel decoders for segmentation and landmark detection, leveraging task synergy.
• Incorporation of a volume of interest (VOI) module and squeeze-and-excitation (SE) blocks with attention mechanisms to enhance feature learning.

Main Results:

• Achieved a mean Dice score of 0.95 for aortic segmentation.
• Attained an average symmetric surface distance of 0.53 mm and a Hausdorff distance of 2.13 mm for segmentation.
• Reached a mean square error (MSE) of 3.23 mm for landmark localization in 40 testing cases.

Conclusions:

• A novel multi-task learning framework effectively performs simultaneous thoracic aorta segmentation and landmark localization.
• The framework demonstrates strong performance, supporting quantitative morphological measurements for analyzing aortic diseases like hypertension.