Self-Supervised Pre-Training with Intensity Guided Masking for Enhanced Aorta Segmentation in CT
Authors:
- 14 July 2025
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Annual International Conference of the IEEE Engineering in Medicine and Biology Society
- Vol. 2025, 1-7
Abstract
Abdominal aortic aneurysm (AAA) is a life-threatening vascular condition that requires regular imaging and follow-ups to prevent fatal outcomes. While accurate diagnosis and selecting treatment strategies depend on aortic segmentation to assess disease progression, manual segmentation is time-consuming, prone to inter-observer variability, and can stall the clinical workflow. For the automatic aorta segmentation several deep learning methods have been proposed with high accuracy. However, their reliance on large annotated databases limits their applicability. To this end, self-supervised learning approaches have been developed to alleviate the need for manual labels during training. In CT imaging, Hounsfield Units (HU) correspond to specific anatomical structures, such as bones and soft tissues, based on their intensity ranges. In this paper, we exploit this property to effectively pre-train a Deep Learning segmentation model using the proposed Intensity Guided Masking (IGM) where we occlude regions within specific intensity ranges in the CT image and aim at predicting/reconstructing the masked area. Next, the pre-trained encoder is integrated into a SwinUNETR model, fine-tuned on manually labeled CT images, and evaluated for aortic structure segmentation. Our proposed method has been evaluated on both a public and a private dataset achieving DSC of 91.20% and 85% and ASSD of 0.05mm and 0.04mm, respectively and outperforming both state-of-the-art supervised baselines and pre-training based methods. The code will be released upon publication at https://github.com/theoVag/SwinUNETR-IGM.Clinical relevance— Our method improves aortic segmentation accuracy in CT imaging while reducing reliance on large annotated datasets, enhancing efficiency in vascular condition assessment such as detecting or quantifying abdominal aortic aneurysms.Keywords
This publication has 27 references indexed in Scilit:
- Kidney Tumor Classification on CT images using Self-supervised LearningComputers in Biology and Medicine, 2024
- SwinMM: Masked Multi-view with Swin Transformers for 3D Medical Image Segmentation2023
- Self-supervised Learning: A Succinct ReviewArchives of Computational Methods in Engineering, 2023
- Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images2022
- AVT: Multicenter aortic vessel tree CTA dataset collection with ground truth segmentation masksData in Brief, 2022
- 3D Automatic Segmentation of Aortic Computed Tomography Angiography Combining Multi-View 2D Convolutional Neural NetworksCardiovascular Engineering and Technology, 2020
- 3D MRI Brain Tumor Segmentation Using Autoencoder Regularization2019
- Tracking Emerges by Colorizing Videos2018
- Medical imaging2015
- Image Analysis for Medical Visualization2014