Abstract
Accurate localization of the seizure onset zone (SOZ) is critical for presurgical epilepsy planning, yet automated channel-level SOZ detection from raw inter-ictal intracranial EEG (iEEG) remains underexplored. Existing outcome-prediction pipelines assume expert SOZ labels as input, creating a circular dependency that motivates end-to-end localization from raw recordings. We propose \emph{Adaptive Riemannian Geometry on Symmetric Positive Definite Neural Manifolds} (ARG-SPD), a geometric deep learning framework for channel-level SOZ candidate detection. ARG-SPD encodes raw iEEG windows with a temporal convolutional neural network, projects features to Symmetric Positive Definite (SPD) covariance matrices on a Riemannian manifold, and classifies them via learnable SPD prototypes with affine-invariant geodesic distances. A Riemannian Focal Loss with spectral dispersion regularization addresses extreme class imbalance ($\sim$8.8\% SOZ channels) and numerical instability. Monte Carlo dropout provides uncertainty-aware triage, separating high-confidence SOZ candidates from ambiguous cases for expert review. On 38,708 interictal iEEG windows from 184 patients (OpenNeuro public dataset) under strict patient-level 5-fold cross-validation, ARG-SPD achieves \textbf{0.82\,$\pm$\,0.03} channel-AUC, outperforming SPDNet by $+$0.06 AUC ($p{<}0.05$) and a 1D-CNN baseline by $+$0.17. High-confidence predictions ($78\%$ of channels) reach $0.89$ AUC, with spatial overlap of $78\%\pm15\%$ Dice against resection zones in a 47-patient subset. These results support ARG-SPD as a promising expert-in-the-loop decision-support tool for presurgical SOZ mapping, pending prospective independent validation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_040.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/profile?id=~Ghazaleh_Khodabandelou1
BibTex
@InProceedings{KhoGha_Adaptive_MICCAISAT2026,
author = { Khodabandelou, Ghazaleh AND Karakas, Cemal AND Linguraru, Marius George AND Anwar, Syed Muhammad},
title = { { Adaptive Riemannian Geometry for Seizure Onset Zone Localization in Intracranial Electroencephalography } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}