Source-linked AI summary
A Generalizable Feature Extractor for Alzheimer's-Related Brain MRI Tasks
Reza Rajabli, D. Louis Collins
TL;DR
Limited labeled data and uncertain cross-dataset transfer motivate this study of reusable neuroimaging representations. The authors freeze a compact brain-age 3D CNN and adapt it with LoRA, finding performance across diagnosis, prognosis, amyloid prediction, and morphometry, including transfer to an unseen cohort without retraining.
Problem
Evidence remains limited on whether transferred neuroimaging models can generalize to new datasets without retraining, especially for Alzheimer’s disease research.
Method
The study freezes a 7.18-million-parameter 3D CNN trained with brain-age supervision and adapts it to six tasks using LoRA with approximately 1% additional trainable parameters.
Results
Across six experiments, the backbone supported diagnosis, prognosis, amyloid prediction, and morphometry, including AUC 0.964 on held-out ADNI folds and AUC 0.871 on OASIS-3 without retraining.
Takeaways & Limitations
A compact brain-age model can serve as a reusable foundation model for Alzheimer’s-related tasks under strict data constraints.
Takeaways & Limitations
The external-transfer results represent a conservative lower bound because hyperparameters were not tuned for each target cohort.
Abstract
from arXiv · showhide
When there is not enough labeled data to properly train deep learning models, transfer learning can help. We still do not fully understand how effective it is in neuroimaging, especially for Alzheimer's disease research. It is also not clear if these transferred models can work on new datasets without being retrained for each specific task. We evaluate whether a compact, supervised pretrained model can serve as a reusable foundation model for downstream neuroimaging tasks. We freeze the 7.18 million weights of a 3D CNN previously trained for brain-age prediction, and adapt it to each task using Low-Rank Adaptation (LoRA), requiring only ~1% additional trainable parameters. We evaluate generalizability in six experiments. Adapting the model to classify cognitively normal versus Dementia on ADNI gave an AUC of 0.964 on held-out folds (Experiment #1). Applying that adapted model unchanged to OASIS-3, with no retraining, gave an AUC of 0.871 (Experiment #2). Reusing its output logit together with age and a cognitive score distinguished stable from progressing MCI with an AUC of 0.828 (Experiment #3). Adapting the same backbone to predict amyloid positivity from structural MRI gave an AUC of 0.804 (Experiment #4). Finally, the same approach estimated ICV-normalized hippocampal and white matter hypointensity volumes directly from the T1w image, with R^2 of 0.80 and 0.91 respectively, tasks normally addressed with much larger U-Net networks (Experiments #5 and #6). A compact model supervised on brain age can therefore serve as a reusable backbone, adapting to each task with ~1% additional parameters and transferring to an unseen cohort without any training. Our findings suggest that a carefully trained brain age model can serve as an effective foundation model for Alzheimer's related tasks, even under strict data constraints.
2 Department of Biomedical Engineering, McGill University
The compact frozen brain-age backbone transfers across Alzheimer’s-related classification, pathology, and morphometry tasks using approximately 1% additional parameters. It estimates structural volumes directly from T1w images and transfers to an unseen cohort without retraining.
- AUC 0.804 was achieved for amyloid-positivity prediction from structural MRI.
- R² of 0.80 and 0.91 estimated ICV-normalized hippocampal and white matter hypointensity volumes, respectively, directly from T1w images.The tasks were performed without explicit segmentation and are normally addressed with much larger U-Net networks.
- The reusable backbone spans diagnosis, prognosis, molecular pathology, and morphometry with approximately 1% additional parameters per task.
- The approach transfers to an unseen cohort without any training on that cohort.
- The findings suggest that a carefully trained brain-age model can serve as an effective foundation model for Alzheimer’s-related tasks under strict data constraints.
Introduction
The introduction frames a gap in neuroimaging transfer learning: age-supervised models have performed competitively but have not been tested as reusable backbones across tasks or on unseen cohorts without retraining. The paper addresses this with a compact frozen 3D CNN adapted using LoRA across six experiments.
- Deep learning for brain MRI typically requires large labeled datasets, while Alzheimer’s research often has small labeled datasets.
- Existing transfer evaluations commonly fit a new classifier on each target dataset, so they do not test performance on entirely unseen data.
- Prior foundation-model studies used contrastive or self-supervised pretraining, whereas age-supervised baselines performed comparably but were not developed beyond classification or target-dataset fitting.
- The proposed model is a compact 3D CNN with 7.18 million parameters whose weights remain frozen while LoRA adds roughly 1% trainable parameters for each task.
- The six experiments cover ADNI diagnosis, unretrained OASIS-3 transfer, MCI prognosis, amyloid prediction, and direct estimation of normalized hippocampal and white matter hypointensity volumes.
- The approach applies the frozen backbone directly to morphometric tasks without explicit segmentation, including normalized hippocampal and white matter hypointensity volumes.
- Freezing the encoder is motivated by evidence that full fine-tuning can improve in-distribution accuracy but reduce out-of-distribution accuracy, while parameter-efficient methods gain advantage as downstream data decreases.
- The compact model uses 7.18 million frozen parameters plus 72,249 parameters per task adapter, supporting deployment on more modest hardware.
Dataset Descriptions
The study uses ADNI for model adaptation and internal evaluation, while OASIS-3 is reserved as an external test cohort. A brain-age pretrained 3D MRI encoder is adapted across tasks with frozen weights and small independently trained LoRA modules.
- Datasets: ADNI is a multi-site longitudinal partnership providing the MRI data used for model development and evaluation.
- Datasets: OASIS-3 contains longitudinal data from over 1,000 subjects and contributes only to external testing, not training or adaptation.
- Data flow: The encoder was pretrained for brain-age prediction, with OASIS-3 deliberately excluded so it remained unseen during pretraining and adaptation.
- Model adaptation: The 7.18-million-parameter base encoder remained frozen while each task received an independently trained LoRA adapter and task head.
- Model adaptation: LoRA added 72,249 trainable parameters, approximately 1% of the model, while the architecture and one hyperparameter setting were retained across experiments.
- Evaluation: Participant-level five-fold cross-validation produced 15 models per experiment, with each model evaluated on its held-out participants.
Experiment #4
Experiment #4 assessed amyloid positivity from structural MRI alone, achieving above-chance discrimination despite considerable overlap between amyloid-positive and amyloid-negative scores. Training used the checkpoint with minimum validation loss to avoid later loss divergence entering the reported results.
- 0.804 AUC was achieved for amyloid positivity from structural MRI alone across 865 subjects, including 405 amyloid-positive subjects.
- 0.734 balanced accuracy, 0.738 sensitivity, and 0.730 specificity were obtained at the 0.5 decision threshold.
- 299 of 405 amyloid-positive subjects were detected, while 124 false positives occurred among 460 amyloid-negative subjects.
- Predicted probabilities were reasonably calibrated, but substantial score overlap reflected the difficulty of inferring amyloid status from structural MRI.
- Validation loss reached its minimum near epoch 4, and the minimum-loss checkpoint excluded later training divergence from reported results.
Experiment #5
Experiment #5 predicted ICV-normalized hippocampal volume directly from MR images, producing close agreement with FreeSurfer measurements. Performance was stable across folds and showed no clear residual trend across the measurement range.
- 0.797 R² was obtained for ICV-normalized hippocampal volume across 2,092 subjects, with a 95% CI of 0.769–0.821.
- 2.6 × 10⁻⁴ mean absolute error accompanied the hippocampal-volume predictions.
- Residuals were close to zero without a clear pattern across measured volumes, indicating no evident systematic prediction bias.
- Validation R² plateaued as loss decreased, and fold-specific R² values ranged from 0.73 to 0.84.
Experiment #6
Experiment #6 predicted ICV-normalized white matter hypointensity volume from the same MR images, achieving strong agreement with measured values. Residuals and training behavior indicated stable performance across the evaluation runs.
- 0.913 R² was achieved for ICV-normalized white matter hypointensity volume across 2,092 subjects, with a 95% CI of 0.893–0.931.
- 6.0 × 10⁻⁴ mean absolute error accompanied the white matter hypointensity predictions.
- Fold-specific R² values ranged from 0.90 to 0.92.
- Residuals were close to zero without a pattern across the predicted range, while bootstrap R² values were narrowly distributed.
- Validation R² plateaued as loss values decreased in all 15 runs.
Discussion
Across six experiments, a compact brain-age encoder adapted with fewer than 1% new parameters performed across Alzheimer’s-related classification and measurement tasks, including transfer to an unseen cohort without retraining. The discussion supports reuse while stressing dataset differences, task-specific limitations, and uncertainty about broader disease generalization.
- Reusable backbone: Fewer than 1% new trainable parameters let the frozen brain-age model support six Alzheimer’s-related experiments, with performance comparable to larger models.The authors frame this as evidence for a reusable feature extractor under data and model-capacity constraints.
- Cross-cohort transfer: AUC 0.964 on held-out ADNI folds and AUC 0.871 on OASIS-3 without retraining demonstrate transfer from within-cohort evaluation to an unseen cohort.The OASIS-3 result used the adapted model unchanged, with no retraining, recalibration, or adaptation.
- Cross-cohort transfer: The external-cohort performance cannot be directly ranked against prior studies because cohorts, preprocessing, and diagnostic definitions differ.The discussion notes that age and selection differences between ADNI and OASIS-3 may also affect the observed performance decrease.
- Scope and limitations: The findings suggest that brain-age pretraining preserves both global and local structural information, but do not establish transfer to diseases with different structural signatures.The authors also note that the white matter hypointensity R² is influenced by a small number of high-burden participants and that hyperparameters were not tuned.
Conclusions
The results support using a compact, brain-age-supervised model as a reusable base for Alzheimer’s-related tasks under strict data constraints. The approach keeps the encoder frozen, adds only a small number of parameters, and can transfer to a new cohort without target training.
- A compact model trained with supervision on brain age is a usable base model for Alzheimer’s-related tasks.
- The encoder remains frozen while each task uses a handful of new parameters and small target datasets.
- The approach also supports a case with no target data at all.
- Including the different LoRA weights, the model has fewer than 7.5 million parameters.
- Existing works contain age-supervised baselines that perform well, but neither pursues them as reusable foundation models.
- The authors suggest this approach can also be used when training a brain foundation model.