Abstract
Non-contrast CT (NCCT) is the primary imaging modality for acute stroke in low- and middle-income countries (LMICs). However, in our West African clinical cohort of 316 scans, only 29 carry expert annotations, proving that specialist labeling, not imaging access, is the bottleneck for AI development. We present the first study applying conditional diffusion models to West African stroke data, investigating annotation efficiency using MedSegDiff-V2 across clinical budgets of N ∈ {5, 10, 15, 20, 29} patients. To balance performance and deployment constraints, we adopt a 2.5D conditioning scheme under fixed training compute. Uniquely, we treat the model’s native stochastic K-pass sampling as a core evaluation axis, assessing Expected Calibration Error (ECE) and predictive entropy alongside standard overlap metrics. Reporting on a pilot subset (Fold 0, K = 3, 2.5D), we analyze high estimator variance and poor calibration. We frame this miscalibration not as a failure, but as a crucial clinical safety signal, arguing that knowing when a model is uncertain is vital in settings lacking routine specialist review.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_045.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=6qwo59jik3
BibTex
@InProceedings{NtiJos_AnnotationEfficient_MICCAISAT2026,
author = { Nti Koduah, Joseph D. A. AND Adjei, Prince Ebenezer AND Akpaloo, Belinda Esinam AND Otoo, Kojo Obed AND Ayivor, Livingstone Eli AND Amuasi, John},
title = { { Annotation-Efficient Conditional Diffusion for Subacute Stroke Segmentation in West African Non-Contrast CT: A Calibration-First Evaluation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}