Do Brain MRI Foundation Models Encode Epilepsy-Relevant Concepts? A Confound-Aware Linear Probing Study on Focal Cortical Dysplasia

Foundation models for brain MRI (e.g., BrainIAC; Tak et al., 2026) are pretrained on broad, heterogeneous cohorts without task-specific supervision for any one pathology. Linear probing asks whether clinically meaningful concepts are nonetheless encoded as linearly decodable directions in such frozen representations. We apply it to focal cortical dysplasia (FCD) type II, a surgically treatable cause of drug-resistant epilepsy where neuroradiologists detect only 64% of lesions on the same 85 public FCD cases used here (Walger et al., 2025). Do frozen embeddings contain linearly decodable information about (a) the presence of an FCD lesion and (b) its hemispheric lateralization, without any fine-tuning on epilepsy data? We extract frozen CLS-token embeddings from BrainIAC (SimCLR-pretrained ViT-B, 768-dim) for all 170 subjects (85 FCD-II patients, 85 healthy controls) of the open Bonn FCD-II MRI dataset (ds004199, OpenNeuro, CC0; Schuch et al., 2023). Each subject contributes two single-channel volumes (one T1w, one FLAIR), each passed separately through the encoder, for $N=340$ scans. We train linear probes (logistic regression) to predict lesion presence and, on the FCD-only subset, hemispheric lateralization, using 5-fold subject-level cross-validation (all scans of a subject in the same fold). Critically, we first run a confound check: a probe predicting acquisition era/scanner protocol from the same embeddings. The Bonn cohort spans two acquisition eras, and MRI encodes site information that models exploit as shortcuts and that intensity harmonization does not remove (Souza et al., 2023). We then report an explicit confound-only baseline against which every clinical probe must be read. The confound probe achieves AUC $0.986 \pm 0.010$: era is almost perfectly decodable and is confounded with the clinical label (163/170 control scans share one protocol). Age points the same way: 23/85 patients fall in the $\leq$20-year brackets vs. 0/85 controls, and age alone reaches AUC $0.622$. Chance is therefore an insufficient reference point, since the question is whether the image embedding carries predictive information beyond such confounds. A trivial classifier using only the one-bit era tag, with no image at all, already separates FCD from control at AUC $0.723$ (CI $[0.668, 0.778]$), a conservative floor given that a binary predictor cannot trade off threshold. The raw lesion-presence probe (FCD vs. control) achieves AUC $0.694 \pm 0.028$ (CI $[0.631, 0.760]$), giving no evidence of incremental predictive value over that era-only baseline. Lateralization ($N=170$ scans / 85 FCD subjects), a task the era tag cannot help with, stays at or below chance both before (AUC $0.436 \pm 0.062$, CI $[0.362, 0.545]$) and after era control ($0.495 \pm 0.083$, CI $[0.391, 0.588]$); flip augmentation during SimCLR pretraining may itself enforce left-right invariance, an independent limit here. As a sensitivity analysis, we remove the acquisition-era mean from each embedding dimension using training-fold statistics only ($N=332$ scans, after dropping 8 rare-tag scans), a lightweight control rather than full harmonization (cf. Fortin et al., 2018); era-controlled lesion-presence gives AUC $0.674 \pm 0.070$ (CI $[0.590, 0.714]$), again no gain over the confound-only baseline. Reported $\pm$ is the standard deviation across folds; all intervals are subject-clustered bootstrap 95% CIs, since each subject contributes two correlated scans. Our key finding is that apparent clinical decodability is explained by readily available confounds: no probe separates itself from a one-bit metadata baseline, and lateralization shows no evidence of a linearly decodable signal, so we cannot distinguish FCD-specific representation from acquisition and demographic confounding in this cohort. Read against chance, the same $0.694$ would have looked like a positive finding, which is precisely our point. Our results suggest that a confound probe plus an explicit confound-only baseline should accompany clinical concept probing of neuroimaging foundation models. More broadly, they illustrate how a dataset can be the best available for a disease and still be unable to support the claim that a foundation model has "already learned" it. References (full citations in the PDF): Tak et al. (2026), Nature Neuroscience 29, 945-956. Walger et al. (2025), Epilepsia Open 10, 778-786. Schuch et al. (2023), Scientific Data 10, 475. Fortin et al. (2018), NeuroImage 167, 104-120. Souza et al. (2023), JAMIA 30(12), 1925-1933.