Abstract
Automated ASPECTS (Alberta Stroke Program Early CT Score) estimation from non-contrast CT (NCCT) remains challenging due to subtle early ischemic changes, low lesion contrast, and inter-reader variability. We propose a clinically motivated BG–SG coupled framework for NCCT-based ASPECTS prediction. Paired basal-ganglia (BG) and supra-ganglionic (SG) slices are processed by independent frozen DINOv3 encoders with LoRA adaptation and separate DPT-style decoders. The two branches do not share weights; cross-level anatomical consistency is enforced only during training, through the proposed Territory-Aware Gated Loss (TAGL), which regularizes corresponding territory pairs M1–M4, M2–M5, and M3–M6. On the public AISD benchmark, the proposed method achieves Dice~=~0.6385 and IoU~=~0.6231 for infarct segmentation. On the ASPECTS-NCCT cohort of 202 scans, the full framework achieves Dice~=~0.772, IoU~=~0.768, ICC~=~0.906, MAE~=~0.684, and Cohen’s~$\kappa$~=~0.884. Additional evaluation on ISLES’22 yields Dice~=~0.699 and IoU~=~0.626. These results indicate that parameter-efficient foundation-model adaptation, combined with anatomically informed supervision, can support clinically meaningful NCCT-based ASPECTS prediction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SWITCH_012.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=KyMl5CB8iP
BibTex
@InProceedings{AzeHib_TerritoryCoupled_MICCAISAT2026,
author = { Azeem, Hiba AND Khan, Behraj AND Ahmad, Shabir AND Syed, Tahir Qasim},
title = { { Territory-Coupled Foundation Model Adaptation for ASPECTS Scoring on Non-Contrast CT } },
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}
}