Abstract
Predicting how a patient’s brain will change over time is a central goal in Alzheimer’s disease (AD), with applications from trial enrichment to patient counselling. Conditional generative models are the natural tool, but face a silent failure mode: consecutive scans of the same subject are nearly identical, so the learning target collapses toward the identity map and the model copies the baseline, undetected by standard image-quality metrics (SSIM, PSNR, FID). We pose longitudinal AD-MRI prediction as stochastic transport problem between a baseline and a follow up distribution and solve it with an Image-to-Image Schrödinger Bridge (I2SB) in the latent space of a biomarker-preserving VAE, conditioned on demographics, APOE4, cognitive scores, treatment, and the time horizon ∆t; this way, the bridge drift stays non-trivial even when endpoints coincide. We add three anti-identity supervision terms to the loss and a suitable evaluation protocol: an identity floor plus a conditional-response test, that measures collapse directly. On held-out subjects the model beats the floor on all three biomarkers, including the hippocampus (the hardest, smallest-signal case), and its follow-ups respond in the clinically correct direction, monotonically in ∆t. The model thus recovers progression signal that the identity baseline cannot: it gets the direction and ordering of change right, while predicting its magnitude conservatively.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_013.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=~Leonida_Lumburovska1
BibTex
@InProceedings{LumLeo_Recovering_MICCAISAT2026,
author = { Lumburovska, Leonida AND Moreira, Pedro AND Bogojeska, Jasmina},
title = { { Recovering Progression Beyond the Identity Shortcut: A Schrödinger Bridge Framework for Longitudinal Brain-MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}