Abstract
Deep learning models for stroke lesion segmentation are commonly evaluated with overlap-based metrics, yet these metrics provide limited insight into whether a model has learned clinically meaningful and data-consistent representations. This limitation is particularly relevant for CT imaging, where lesions are less visible and annotations may be transferred from MRI, introducing spatial uncertainty and label noise. We propose an area-guided latent space analysis framework for segmentation quality control. Three U-Net variants were trained on CT and FLAIR MRI stroke images: a baseline U-Net, an output-constrained U-Net with an area loss on the predicted mask, and a feature-supervised U-Net with an auxiliary lesion-area regression branch at the bottleneck. Bottleneck features were projected into a two-dimensional PCA space, and their monotonic alignment with lesion area was evaluated using Spearman correlation. Feature-level area supervision produced the strongest latent-area alignment, reaching correlations of 0.655 in CT and 0.806 in FLAIR MRI. Quality-control analysis was then performed using patient-wise out-of-fold predictions and latent representations. For each slice, a post-hoc incoherence score compared its predicted lesion area with the predicted areas of its nearest latent neighbors. The score therefore required neither a reference mask nor dedicated quality-control labels at inference time. In CT, incoherent slices had a DSC of 0.08 compared with 0.37 for coherent slices, while in FLAIR MRI the corresponding values were 0.17 and 0.64. Incoherent slices also showed larger boundary and lesion-area errors and were strongly enriched in small lesions. These results indicate that weak global supervision of the bottleneck can provide a reference-free post-hoc signal for prioritizing atypical or unreliable segmentations for manual review.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SWITCH_024.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=5oqsiJ8JhN
BibTex
@InProceedings{MorJul_Areaguided_MICCAISAT2026,
author = { Moreau, Juliette AND Mechtouff, Laura AND Rousseau, David AND Berthezène, Yves AND Eker, Omer AND Cho, Tae-Hee AND Frindel, Carole},
title = { { Area-guided latent coherence for post-hoc segmentation quality assessment in stroke imaging } },
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}
}