A Data-Centric Approach for Segmenting the Aortic Vessel Tree: A Solution to SEG.A. Challenge 2023 Segmentation Task

Data-centric AI is a discipline that focuses on improving the quality and relevance of data used to train AI models. It is a shift from the traditional model-centric approach, which focuses on improving the performance of AI models by tuning the model parameters. This paper presents a data-centric approach for segmenting the aortic vessel tree. The proposed approach consists of a preprocessing pipeline that performs histogram matching and sigmoid windowing, followed by a series of 3D segmentation models. The preprocessing pipeline is designed to improve the contrast and visibility of the vessels in the images, which makes the task easier for the segmentation models. There are three stages of UNet networks each of which performs a different level of segmentation where the result progresses from coarse to fine segmentation. We trained and evaluated the approach to the segmentation of the aorta challenge SEG.A. 2023 dataset. Our approach achieved a Dice Similarity Score of 0.92  \( \pm \)  0.02 and a Hausdorff Distance (95%) of 6.3 mm  \( \pm \)  5.72. Our approach produced a segmentation pipeline that accurately captures the complex structure of the aortic vessel tree and is resistant to changes in noise level, contrast, and geometry.