Forecasting how neurodegeneration reshapes the brain is central to earlier intervention, sharper clinical trials and more tailored care. Magnetic Resonance Imaging (MRI) captures structural change, yet many artificial intelligence approaches struggle to personalise predictions, exploit longitudinal information, maintain temporal coherence and process full 3D scans efficiently. Brain Latent Progression (BrLP) addresses these barriers with a spatiotemporal framework that predicts individual future brain MRIs from 3D data. It pairs a diffusion-based generator with prior disease knowledge, conditions outputs on subject information and introduces a stabilisation step that improves consistency over time while exposing prediction uncertainty. Trained on 11,730 T1-weighted MRIs from 2,805 subjects and validated on an external set of 2,257 MRIs from 962 subjects, BrLP reports accuracy gains over established methods.
Addressing Core Barriers to Personalised Forecasting
BrLP targets four linked challenges observed in existing approaches. First, individualisation: predictions are conditioned on subject metadata, including age, sex and cognitive status, together with region-specific volumetric cues associated with Alzheimer’s disease (AD). This moves beyond population averages and scalar biomarker trends to reflect individual anatomy and likely change. Second, longitudinal exploitation: an auxiliary progression model injects prior knowledge of how AD-relevant regions evolve over time, enabling the system to use information from previous visits when available and to estimate target-age volumes otherwise. Third, spatiotemporal consistency: inference can produce small stochastic variations between successive time points, so BrLP introduces Latent Average Stabilisation (LAS), which averages multiple generations to smooth trajectories and reduce irregularities. Fourth, memory demand: the framework operates in a compact latent space, allowing end-to-end 3D synthesis without reverting to slice-based reconstructions or low-resolution volumes that risk losing clinically relevant detail.
These design choices respond to limitations noted across prior models. Slice-wise pipelines can miss inter-slice dependencies that matter for 3D disease dynamics. Deformation-only methods struggle to create new structures if they are not present at baseline. Models that ignore metadata may forgo useful signals for personalisation. By integrating metadata, longitudinal priors and latent-space diffusion, BrLP seeks to retain anatomical fidelity while remaining computationally tractable.
Must Read: Self-supervised Alignment Flags Abnormal Brain MRI
How the BrLP Framework Works
The framework couples four components. A Latent Diffusion Model learns the distribution of brain MRIs in a compact representation and supports conditional generation. A ControlNet steers synthesis towards the individual’s baseline anatomy by conditioning on the subject’s encoded scan. An auxiliary model contributes disease-informed targets for key regions linked to AD progression, including hippocampus, amygdala, cerebral cortex, cerebral white matter and lateral ventricles. These predicted volumes, together with subject metadata at the target age, form the conditioning used during generation. LAS repeats inference several times, averages the resulting latents, then decodes a single future MRI, improving temporal smoothness and enabling uncertainty estimation.
Uncertainty is quantified globally from the spread of latent predictions and voxel-wise by computing variance across the decoded images, yielding an uncertainty map that highlights less reliable areas. In ablation, increasing the LAS repetition parameter improved image similarity and reduced volumetric errors, with a trade-off in computation. The auxiliary model and LAS each contributed gains on their own and together delivered the strongest performance. In the ablation setting that split each subject’s timeline into input and prediction halves, adding the auxiliary model reduced volumetric error in conditioned regions by 23%, LAS added a further 4% reduction, and the combination achieved a 21% reduction across evaluated regions compared with the base configuration without these components.
Validation and Comparative Performance
Training and internal evaluation used three longitudinal cohorts pooled into a single dataset, each subject contributing at least two visits. External validation used a separate longitudinal cohort that was not seen during training. Preprocessing included bias-field correction, skull stripping, registration to a common space, intensity normalisation and resampling to isotropic resolution. Performance was assessed by comparing generated MRIs to real follow-up scans using image similarity and volumetric error in disease-relevant regions, with cerebrospinal fluid and thalamus left unconditioned to test generalisation beyond directly modelled covariates.
Against established baselines covering single-image and sequence-aware approaches, BrLP consistently produced lower error and higher similarity on both internal and external test sets. Reported summaries show substantial reductions in mean squared error and increases in structural similarity relative to comparators across subject groups stratified by cognitive status. Visual comparisons illustrated closer alignment with observed anatomical change, including enlargement of lateral ventricles and cortical alterations over time. Additional experiments examined sensitivity to cognitive-status conditioning. When Alzheimer’s cases were conditioned as cognitively normal, volumetric errors increased, particularly in the hippocampus. In sequence-aware settings, access to longitudinal history via the auxiliary model mitigated part of this effect by informing progression rates from earlier visits.
BrLP advances individualised forecasting of brain structural change by uniting latent diffusion, disease-informed conditioning and inference-time stabilisation with quantified uncertainty. The framework processes full 3D MRIs in latent space, incorporates subject metadata and predicted regional volumes, and improves temporal coherence through averaging. Trained on a large longitudinal collection and validated externally, it reports gains over established baselines while offering practical tools for uncertainty assessment. These capabilities point to a route for tracking neurodegeneration at the level of the individual, supporting patient monitoring, trial enrichment and longitudinal research without sacrificing spatial detail or computational feasibility.
Source: Medical Image Analysis
Image Credit: iStock
References:
Puglisi L, Alexander DC & Ravì D (2025) Brain Latent Progression: Individual-based spatiotemporal disease progression on 3D Brain MRIs via latent diffusion. Medical Image Analysis; 106: 103734.
