Deep Learning for Vessel Segmentation in CTAs
Keywords: Image segmentation, Image labeling, Deep Learning, Medical imaging
TL;DR: Deep learning for vessel segementation across multiple anatomies using CTAs
Abstract: Accurate vessel segmentation from Computed Tomography Angiography (CTA) is crucial for clinical analysis, but manual methods are laborious, motivating the need for automated solutions. This work presents and evaluates a deep learning methodology for automated vessel segmentation across multiple anatomical regions in CTAs. Using a custom-developed labeling tool, we created a dataset of ~54,000 patches from 51 CTAs. We employed a multitask learning Attention U-Net model, which combined vessel segmentation with body part classification, achieving an average classification F1 score of 0.866 and segmentation dice score of 0.87 across chest, abdomen, pelvis, and thigh with 5-fold cross-validation. Our findings offer insights into building effective automated vessel segmentation models, and we contribute our open-source labeling tool and training code to facilitate further research. Github: https://github.com/Ross-Lab-UCSD/cta-vessel-segmentation
Submission Number: 86
Loading