A model fusion method based DAT-DenseNet for classification and diagnosis of aortic dissection

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Cardiovascular Diagnostics

Background:

• Accurate diagnosis of aortic dissection is critical for patient outcomes.
• Current diagnostic methods rely on interpreting complex imaging data, presenting challenges.
• Automated analysis of CT angiography (CTA) images can aid in diagnosis.

Purpose of the Study:

• To develop a robust method for accurate patient-level aortic dissection diagnosis using CTA images.
• To propose and evaluate a novel deep learning model, DAT-DenseNet, for this task.
• To introduce a feature fusion module for improved patient-level classification.

Main Methods:

• A classification model, DAT-DenseNet, combining deep attention Transformer and DenseNet architectures was developed.
• Two DAT-DenseNet models were used in parallel for image-level classification tasks.
• A feature fusion module was proposed to combine image features for patient-level classification.

Main Results:

• DAT-DenseNet achieved 92.41% accuracy at the image level, surpassing common models by 2.20%.
• The proposed feature fusion method resulted in 90.83% accuracy at the patient level.
• Experimental results confirmed the reliability and high performance of the DAT-DenseNet model and fusion strategy.

Conclusions:

• DAT-DenseNet demonstrates high performance for image-level classification of aortic dissection.
• The feature fusion module effectively maps image features to patient outcomes, enabling accurate patient classification.
• The proposed method offers a reliable approach for computer-aided diagnosis of aortic dissection.