Automatic Detection and Segmentation of Thrombi in Abdominal Aortic Aneurysms Using a Mask Region-Based Convolutional...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease

Background:

• Thrombus detection and segmentation are critical for managing abdominal aortic aneurysms (AAAs).
• Deep convolutional neural networks (CNNs) show promise for improving these tasks, but current methods often focus solely on segmentation after detection.
• Existing CNN approaches for thrombus analysis are still in early stages of investigation.

Purpose of the Study:

• To develop a fully automated method for both thrombus detection and segmentation in AAAs.
• To enhance the Mask R-CNN framework with optimized loss functions for improved performance.

Main Methods:

• A Mask R-CNN framework was adapted and improved with optimized loss functions.
• Complete intersection over union (CIoU) and smooth L1 loss were used for accurate thrombus detection.
• Modified focal loss was employed to enhance thrombus segmentation accuracy.

Main Results:

• The method was evaluated on 60 computed tomography angiography (CTA) patient studies using 4-fold cross-validation.
• Achieved the highest F1 score for thrombus detection (0.9197) compared to state-of-the-art methods.
• Demonstrated superior performance across most metrics for thrombus segmentation.

Conclusions:

• The proposed fully automated method offers superior performance for thrombus detection and segmentation in AAAs.
• This approach advances the application of deep learning in cardiovascular imaging and patient management.
• The optimized loss functions contribute to enhanced accuracy in analyzing AAA thrombi.