Abstract
Transfer learning from healthy to tumor-bearing functional magnetic resonance imaging (fMRI) improves classification of resting-state networks (RSNs), yet the same models lose a few accuracy points on the unhealthy class. We ask which image-level property drives that residual gap. Using $81$ healthy controls and $55$ patients with brain tumors, we train a convolutional neural network (CNN) to discriminate the two classes ($0.95 \pm 0.02$ accuracy) and inspect its decisions with gradient-weighted class activation mapping (Grad-CAM) and a t-distributed stochastic neighbor embedding (T-SNE) of the latent space. We then run four statistical tests on the image content: lesion versus perilesional intensity, the intersection-over-union (IoU) between RSN activation and the tumor mask, activation intensity with versus without tumor overlap, and per-subject activated volume. None of the lesion-local tests survives a sham-mask test, in which the lesion mask is mirrored into tumor-free tissue, whereas activated volume separates the cohorts ($p = 6.5 \times 10^{-15}$, rank-biserial $r = -0.79$), and this holds in all seven RSNs and after accounting for lesion size. The discriminative signal therefore appears not to be local to the lesion but a global, network-wide reduction in activated volume, consistent with tumor-induced suppression of the blood-oxygen-level-dependent (BOLD) response, though our data identify an association rather than a cause. This reframes the validation drop as a global statistical shift rather than a local lesion artifact, and suggests that uniform z-thresholding across cohorts is inappropriate when activation density itself differs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_064.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=w46WctYmnv
BibTex
@InProceedings{IsmLuk_Where_MICCAISAT2026,
author = { Ismaila, Lukman E. AND Rasti, Pejman AND Lemée, Jean-Michel AND Rousseau, David},
title = { { Where Does the Discrepancy Come From? Probing Healthy vs. Tumor fMRI with Convolutional and Statistical Lenses } },
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}
}
