An innovative bimodal computed tomography data-driven deep learning model for predicting aortic dissection: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diseases

Background:

• Aortic dissection (AD) is a life-threatening condition requiring rapid diagnosis.
• Current computed tomography angiography (CTA) diagnosis is time-consuming due to contrast agents.
• Existing deep learning (DL) models are limited to single-modality inputs (non-contrast CT or CTA).

Purpose of the Study:

• To develop a bimodal deep learning (DL) framework for improved aortic dissection (AD) diagnosis.
• To enable independent processing of non-contrast CT and CTA data for dual-path detection.
• To enhance the efficiency of AD diagnosis.

Main Methods:

• A retrospective study included patients who underwent non-contrast CT and CTA between February 2016 and September 2021.
• A two-stage DL model was developed: an aorta detection network (AoDN) followed by an aortic dissection diagnosis network (ADDiN).
• Performance was assessed using average precision at the intersection over union threshold 0.5 (AP@0.5) for aorta detection and area under the receiver operating characteristic curve (AUC) for diagnosis.

Main Results:

• The AoDN achieved high performance in aorta detection, with AP@0.5 of 99.14% on non-contrast CT and 99.34% on CTA.
• The ADDiN demonstrated excellent diagnostic accuracy, achieving AUCs of 0.98 for non-contrast CT and 0.99 for CTA.
• The model was validated across three institutions.

Conclusions:

• The proposed bimodal CT data-driven DL model accurately diagnoses AD.
• This framework facilitates prompt hospital diagnosis and treatment initiation for AD patients.
• The study highlights the potential of multimodal DL in improving cardiovascular emergency diagnostics.