Career Center AES Connect

Abstracts

Introducing Automated and Interpretable Detection of Hippocampal Sclerosis (AID-HS) at a District Epilepsy Centre: Observations and Challenges

Abstract number : 2.306
Submission category : 5. Neuro Imaging / 5A. Structural Imaging
Year : 2025
Submission ID : 73
Source : www.aesnet.org
Presentation date : 12/7/2025 12:00:00 AM
Published date :

Authors :
Presenting Author: Matthew Suppa Modrusan, BScH – Queen's University

Aikam Rai, BHS – Queen's University
Andrea Ellsay, MSc – Stanford University
Konrad Wagstyl, MBPhD – King's College London
Mathilde Ripart, PhD – King's College London
Donald Brien, MSc – Queen's University
Donatella Tampieri, MD – Kingston Health Sciences Centre
Ada Mullett, MA – Kingston Health Sciences Centre
Ron Levy, MD – Kingston Health Sciences Centre
Lysa Lomax, MD/MSc – Kingston Health Sciences Centre
Garima Shukla, MD – Kingston Health Sciences Centre
Karla Batista Garcia-Ramo, PhD – Queen's University
Gavin Winston, MD/PhD – Queen's University

Rationale: In patients with focal epilepsy, resective surgery can render many seizure-free. However, hippocampal sclerosis (HS)—a common pathology in temporal lobe epilepsy (TLE)—often eludes visual detection on MRI, causing potential surgical delays. Bilateral HS is particularly difficult to evaluate given the lack of asymmetry. Advanced MRI postprocessing can improve detection but requires real-world validation. We introduced AID-HS (Ripart et al., 2025)—an open-source tool from the Multicentre Epilepsy Lesion Detection (MELD) project—into our presurgical pathway to assess its utility.

Methods: Following initial review in epilepsy surgery rounds, patients with TLE underwent 3T MRI with a Harmonized Neuroimaging of Epilepsy Structural Sequences (HARNESS) protocol. Scans were reviewed by a neuroradiologist and postprocessed with AID-HS, which characterizes hippocampal morphology and asymmetries to produce probability scores for left HS, right HS, or no asymmetry. New findings and concordance with clinical data were recorded for retrospective cases (with histopathology) and prospective cases under review.

Results: In the retrospective post-surgical cohort (n=8), AID-HS correctly detected and lateralized unilateral HS in both confirmed cases. However, it also issued bilateral HS warnings in 3 cases without histopathological evidence of HS (2 mesial temporal astrogliosis, 1 temporal FCD). In the prospective cohort (n=9), AID-HS detected and lateralized unilateral HS in 1 case (concordant with imaging) and flagged potential bilateral HS in 5 cases. In these patients, image review noted hippocampal atrophy but no other features of HS, such as FLAIR hyperintensity. Notably, 3 patients were MRI-negative with suspected bilateral TLE based on clinical and EEG data. Across both cohorts, bilateral HS warnings occurred in patients with 14% lower total brain volume (TBV) (1000 cm³ vs. 1167 cm³) and a 2.6-fold longer duration of epilepsy (14.0 years vs. 5.4 years), suggesting hippocampal volume loss in the context of global brain atrophy. Figure 1 shows generalized atrophy in a patient with suspected bilateral TLE. Figure 2 shows the corresponding output from AID-HS.

Conclusions: AID-HS is helpful for detecting unilateral HS. However, in patients with longstanding epilepsy, global brain atrophy may lead to the incorrect warning of bilateral HS. This warrants caution when other clinical data suggest bilateral TLE. Challenges are compounded when neuroradiological opinion conflicts with an algorithm—particularly in prospective cases where no gold standard is available. Three limitations may explain this issue. First, bilateral HS lacks the asymmetries typically used to guide radiological detection. Second, AID-HS relies solely on T1-weighted MRI, omitting key features visible on T2-weighted/FLAIR scans. Third, AID-HS normalizes hippocampal features to age-matched healthy controls without adjusting for individual TBV. In patients with global brain atrophy, this may lead to spurious warnings of bilateral HS. The MELD team is currently training models that incorporate additional imaging data and correct for TBV across a larger cohort.

Funding: Southeastern Ontario Academic Medical Organization Innovation Fund.

Neuro Imaging