A semi-supervised diffusion model for automated segmentation of vascular plaques
Related Experiment Video
Updated: Oct 31, 2026
A Label-Free Segmentation Approach for Intravital Imaging of Mammary Tumor Microenvironment
Published on: May 24, 2022
Jianfei Liu1, Praveen Thoppey Srinivasan Balamuralikrishna2, Langston Locke2
1Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Clinical Center, National Institutes of Health, 10 Center Dr, Bethesda, Maryland, 20892, USA. jianfei.liu@nih.gov.
Purpose:
Segmenting vascular plaques in the carotid arteries, aorta, and iliac arteries on CT is important because plaque burden relates to cardiovascular and stroke risk and can guide treatment. However, manual labeling is often infeasible and time-consuming across large vascular regions. Semi-supervised learning (SSL) can leverage unlabeled CT data to enable accurate plaque segmentation with only limited labels.
Methods:
We applied a diffusion-based semi-supervised learning (SSL) method for vascular plaque segmentation in non-contrast CT by formulating the task as multi-class segmentation of plaques and 26 adjacent anatomical structures. The method employs a teacher-student architecture with a shared diffusion encoder to learn domain-invariant feature representations and separate teacher and student decoders to decouple labeled and unlabeled data, reducing overfitting and improving robustness. To address the severe class imbalance caused by the small size of plaques, we developed a plaque-aware oversampling strategy and a labeled-data-guided training for the shared encoder and student decoder. We used 142 NIH non-contrast CT scans, divided into 119 training scans and 23 validation scans. We externally evaluated the method on 95 scans from the University of Wisconsin and 50 scans from Medipol University Mega Hospital and compared its performance with supervised nnU-Net and a strong SSL baseline.
Results:
The improved SSL method significantly outperformed the original SSL method and supervised training with the same architecture when using 14 labeled training scans ( ). With these 14 labeled training scans, it achieved DSC values of 58.8% and 82.3% on CTRATE and Renal Donor, respectively, within 2.7--2.8 percentage points of the same architecture trained with all 119 training scans labeled.
Conclusion:
The SSL method segments plaques in the carotid, aortic, and iliac arteries with performance approaching that of fully supervised training while using only 14 labeled training scans, supporting scalable plaque burden assessment.
