# Automatic Explainable Segmentation of Abdominal Aortic Aneurysm From Computed Tomography Angiography

## TL;DR

This paper introduces an automated deep learning framework for accurately segmenting abdominal aortic aneurysms in CT scans, with high precision and real-time processing capabilities.

## Contribution

A novel DL framework with a dynamic router and specialized U-Net models for precise and efficient AAA segmentation in CTA images.

## Key findings

- The model achieved dice scores of 0.9648 for aortic lumen and 0.9615 for outer wall segmentation.
- The system processes each image frame in 17 ± 1 milliseconds, suitable for real-time use.
- NURBS refinement improved accuracy in complex cases with correction times of 3–20 seconds per frame.

## Abstract

This work presents an automated deep learning (DL) based framework for segmenting abdominal aortic aneurysm (AAA) in contrast-enhanced computed tomography angiography (CTA) images, which was developed to support AAA screening and analysis. The framework includes a dynamic router that assigns image regions to three specialized U-Net models, each trained to handle different aspects of the segmentation. It was trained and validated on 9,080 images and tested on 1,560 images representative of 22 unique patients. The model accurately segmented both the aortic lumen and the outer wall, achieving dice scores (DS) of 0.9648 and 0.9615, intersection over union (IoU) scores of 0.9324 and 0.9264, and Hausdorff distance (HD95) percentile values of 1.3490 mm and 1.3670 mm, respectively. The fully automated system processes each image frame in approximately 17 ± 1 milliseconds, making it suitable for real-time use. In certain complex cases where improved clinical accuracy is required, non-uniform rational B splines (NURBS) were used to manually refine the segmentation. In these cases, the NURBS correction time ranges from 3 to 20 seconds per frame. The framework’s training and validation demonstrate its potential as a reliable tool for AAA detection and clinical decision-making. Future work should focus on integrating multimodal imaging and optimization of NURBS to further improve its accuracy and efficiency.