Abstract

Spectral CT angiography with deep learning–based infarct core segmentation may aid in thrombectomy eligibility assessment in acute ischemic stroke. However, several inherent sources of data variability, such as acquisition timing and lack of a reliable ground truth, complicate core segmentation. We therefore developed a data-uncertainty-aware U-Net, trained on iodine density (ID) and non-contrast CT images with aleatoric uncertainty estimation in its loss function, that is robust to uncertain input and noisy training labels. U-Net predictions were compared against relative ID threshold segmentations and CT perfusion (CTP)–derived references, with a radiologist scoring and ranking them across 39 cases. U-Net segmentations significantly outperformed thresholded relative ID on Likert score (p = 3.4e-4, r = 0.60) and on ranking (p = 6.4e-5, r = 0.67). CTP references were only significantly better when the radiologist ranked the segmentations (p = 2.9e-2, r = 0.35). The Likert scores of the CTP references were not significantly different from those of the U-Net segmentations (p = 9.3e-2, r = 0.27). U-Net ranked first 10 out of 39 cases while the reference ranked first 25 out of 39 cases, showing good generalizability of the U-Net and imperfections of the CTP-derived references. Incorporating uncertainty prevented over-segmentation, leading to better volumetric similarity with the CTP references. These results suggest a data-uncertainty-aware U-Net applied to spectral CT angiography ID maps could be a valuable alternative when CTP is unavailable.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SWITCH_015.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=Qj5EDWKF0j

BibTex

@InProceedings{VroJor_Infarct_MICCAISAT2026,
        author = { Vromans, Joris AND Bennink, Edwin AND Dankbaar, Jan Willem AND Velthuis, Birgitta K. AND de Jong, Hugo W. A. M.},
        title = { { Infarct Core Segmentation on Spectral CT Angiography using a Data-Uncertainty-Aware U-Net } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17255},
        month = {pending},
        page = {pending}
}


back to top