3D-Neuro-SimCLR: Brain MRI Model

• The paper introduces a 3D SimCLR-based self-supervised method trained on nearly 45,000 T1 scans from 11 public datasets to learn general brain representations.
• It employs a high-resolution, standardized preprocessing pipeline with TurboPrep to preserve subtle anatomical details across diverse neurological conditions.
• SimCLR fine-tuning consistently outperforms both MAE and supervised baselines in low-label and out-of-distribution settings, enhancing clinical transferability.

Searching arXiv for the specified papers to ground the article in the current preprint record. 3D-Neuro-SimCLR is a public, high-resolution self-supervised foundation model for 3D brain structural MRI built with SimCLR-style contrastive learning on 18,759 patients and 44,958 T1 scans from 11 publicly available datasets (Kaczmarek et al., 12 Sep 2025). It is designed to address a recurrent limitation in brain MRI artificial intelligence: most 3D MRI models are built for one task, one pathology, or one dataset, and they usually depend heavily on labeled data. The model is intended to learn general brain representations from diverse, mostly unlabeled T1-weighted scans spanning multiple neurological diseases and healthy cohorts, and to transfer those representations to downstream clinical prediction tasks in both in-distribution and out-of-distribution settings (Kaczmarek et al., 12 Sep 2025).

1. Scope and problem setting

The central premise of 3D-Neuro-SimCLR is that 3D structural brain MRI has enough anatomical commonality across diseases that a single general-purpose foundation model can be trained on large, diverse, unlabeled data and then reused across multiple downstream clinical problems (Kaczmarek et al., 12 Sep 2025). The work positions itself against three limitations in prior 3D brain MRI foundation models: training on too few datasets or too few pathologies, low-resolution or heavily cropped volumes, and unreleased or hard-to-reproduce pipelines.

The motivation is closely tied to the practical structure of neuroimaging datasets. Labels are expensive, diseases can be rare, and even common labels can be noisy or ambiguous. In that setting, a self-supervised learning framework is intended to reduce dependence on task-specific annotation while preserving transferability across populations, scanners, hospitals, and diagnoses. This suggests an explicit attempt to treat structural neuroanatomy, rather than any single diagnostic label, as the principal source of reusable signal.

A relevant conceptual predecessor is the earlier 3D SimCLR literature in medical imaging, which adapted contrastive learning to volumetric data and showed that self-supervised pretraining can improve downstream data-efficiency in 3D segmentation tasks (Ali et al., 2021). Relative to that earlier line of work, 3D-Neuro-SimCLR targets a broader brain MRI foundation-model setting rather than tumor or organ segmentation alone.

2. Pretraining corpus and dataset composition

The pretraining corpus consists of 18,759 patients, 44,958 total T1 scans, and 11 publicly available datasets (Kaczmarek et al., 12 Sep 2025). The training sources are ADNI, AOMIC, CoRR, DLBS, GSP, HABS-HD, MCSA, NIFD, PPMI, SALD, and SOOP. The datasets include both healthy and diseased cohorts and cover Alzheimer’s disease, mild cognitive impairment, Parkinson’s disease, frontotemporal dementia, stroke, and healthy controls.

The authors use T1-weighted structural MRI from every dataset because it is the most common MRI sequence across the sources and is appropriate for learning general brain anatomy (Kaczmarek et al., 12 Sep 2025). Several datasets also contain mixed or underrepresented conditions beyond their primary diagnosis. HABS-HD includes healthy controls, MCI, AD, and also other disorders like depression, anxiety, and traumatic brain injury; PPMI is mainly Parkinson’s but includes rare neurological disorders too. In the paper’s framing, this heterogeneity is part of what makes the model closer to a general foundation model.

Component	Details
Patients	18,759
Total T1 scans	44,958
Public datasets	11
Imaging sequence	T1-weighted structural MRI
Training sources	ADNI, AOMIC, CoRR, DLBS, GSP, HABS-HD, MCSA, NIFD, PPMI, SALD, SOOP

The dataset design is significant because it explicitly emphasizes disease diversity and healthy controls within a single pretraining pool. A plausible implication is that transfer performance is being driven not only by scale, but also by cross-pathology heterogeneity.

3. Preprocessing and high-resolution representation

The preprocessing pipeline is standardized and public, and it uses TurboPrep with reproducible, fast preprocessing steps: N4 bias-field correction, skull stripping, linear registration to a template, brain segmentation, and intensity normalization (Kaczmarek et al., 12 Sep 2025). Skull stripping and segmentation are performed with trained deep-learning components in TurboPrep, and the whole pipeline takes about one minute per scan on average.

All images are resampled to 1 × 1 × 1 mm³, which the authors treat as high resolution for 3D MRI. This corresponds to volumes of 193 × 229 × 193 before cropping. The images are then processed by permuting axes so depth is first, center-cropping to 150 × 192 × 192, and applying per-sample z-score normalization (Kaczmarek et al., 12 Sep 2025). The paper identifies this high-resolution preprocessing as important for preserving small lesions and subtle anatomy.

This resolution choice distinguishes the model from prior work described as limited by low resolution or heavy cropping. The emphasis on preserving subtle anatomy is particularly relevant in neurological disease settings in which clinically informative variation may be spatially localized or morphometrically small.

4. SimCLR formulation and architecture

The self-supervised objective is 3D SimCLR, adapted from the standard contrastive framework to 3D MRI (Kaczmarek et al., 12 Sep 2025). For each 3D scan $x$, two augmented views are created,

$x_i = t_i(x), \qquad x_j = t_j(x),$

which are passed through the same encoder $f$ to obtain representations $h_i$ and $h_j$, then through a projection head $g$ to obtain $z_i$ and $z_j$.

The contrastive loss is the normalized temperature-scaled cross-entropy loss:

$\mathcal{L}_{i,j} = -\log \frac{\exp(\mathrm{sim}(z_i, z_j) / \tau)}{\sum_{k=1}^{2N} \mathbbm{1}_{[k \ne i]} \exp(\mathrm{sim}(z_i, z_k) / \tau)}$

with cosine similarity

$$\text{sim}(z_i, z_j) = \frac{z_i^\top z_j}{\|z_i\|\|z_j\|}.$$

The model averages both directions $x_i = t_i(x), \qquad x_j = t_j(x),$0 and $x_i = t_i(x), \qquad x_j = t_j(x),$1, then averages over the batch (Kaczmarek et al., 12 Sep 2025). Operationally, the objective pulls together embeddings of two augmented views of the same MRI and pushes apart embeddings of different MRIs in the batch.

The encoder is a 3D CNN based on ResNet-18, chosen because medical datasets are still modest relative to natural-image datasets and CNNs have stronger inductive bias in that setting (Kaczmarek et al., 12 Sep 2025). The projection head has output dimension 64. For SimCLR pretraining, augmentations are adapted to grayscale 3D MRI and implemented with MONAI: random spatial crop with random center and minimum size $x_i = t_i(x), \qquad x_j = t_j(x),$2, resize to $x_i = t_i(x), \qquad x_j = t_j(x),$3, axial flip with $x_i = t_i(x), \qquad x_j = t_j(x),$4, 45° rotation with $x_i = t_i(x), \qquad x_j = t_j(x),$5, intensity shift with $x_i = t_i(x), \qquad x_j = t_j(x),$6 and offset 0.5, and gamma/contrast adjustment with range $x_i = t_i(x), \qquad x_j = t_j(x),$7 and $x_i = t_i(x), \qquad x_j = t_j(x),$8 (Kaczmarek et al., 12 Sep 2025). Training uses 12 H100 GPUs with an effective batch size of 72 for 150 epochs.

The underlying contrastive formulation closely matches earlier 3D SimCLR approaches in medical imaging, which also use volumetric augmentations, a 3D CNN encoder, and the NT-Xent objective (Ali et al., 2021). The principal difference is scope: 3D-Neuro-SimCLR trains directly on full structural brain MRI scans for a foundation-model objective across neurological diseases.

5. Comparative baselines and evaluation design

The paper compares the SimCLR model against a 3D Masked Autoencoder and two supervised baselines (Kaczmarek et al., 12 Sep 2025). In the MAE setup, the volume is divided into non-overlapping patches, 75% of tokens are masked, the encoder sees only the unmasked tokens, and a decoder reconstructs the masked patches in voxel space. The MAE loss is reconstruction MSE on masked patches:

$x_i = t_i(x), \qquad x_j = t_j(x),$9

The MAE backbone is ViT, with three sizes tested: ViT-Tiny, ViT-Base, and ViT-Large. The patch size is 25 × 16 × 16, the total number of patches is 864, positional embeddings are sinusoidal, and training uses 8 H100 GPUs, effective batch size 128, and 150 epochs (Kaczmarek et al., 12 Sep 2025).

The supervised baselines are ResNet-18, corresponding to the SimCLR backbone, and ViT-Tiny, corresponding to the MAE backbone. These are trained from scratch using the same downstream data splits and MONAI augmentations: random crop with minimum size $f$0, resize to $f$1, axial flip with $f$2, affine transforms with rotation range 0.1 rad, scale range 15%, translation range 5 voxels with $f$3, intensity shift with $f$4, and Gaussian noise with $f$5 (Kaczmarek et al., 12 Sep 2025).

Downstream evaluation covers four tasks, including both in-distribution and out-of-distribution settings:

Task	Dataset	Target	Metric
In-distribution regression	SOOP	NIH Stroke Scale score	MAE
OOD classification	AIBL	AD vs healthy controls	AUC
OOD classification	IXI	Sex classification	AUC
OOD regression	IXI	Age prediction	MAE

The SOOP task uses a 60/20/20 split, with the training split used for SSL pretraining and downstream fine-tuning while validation and test are held out. The AIBL task uses only 106 training patients total, balanced across classes, while validation and test remain imbalanced with very few AD cases. IXI contains healthy subjects only, making sex classification and age regression tests of whether the representation captures general brain anatomy rather than only disease cues (Kaczmarek et al., 12 Sep 2025).

6. Quantitative results and low-label behavior

The principal reported result is that fine-tuned SimCLR is best on every task (Kaczmarek et al., 12 Sep 2025). For stroke severity regression on SOOP, the scores are ResNet-18: 5.47 ± 0.09, ViT-T: 6.02 ± 0.09, MAE-LP: 6.12 ± 0.02, MAE-FT: 6.15 ± 0.15, SimCLR-LP: 5.87 ± 0.07, and SimCLR-FT: 5.37 ± 0.24, with lower MAE indicating better performance.

For Alzheimer’s classification on AIBL, the scores are ResNet-18: 0.869 ± 0.030, ViT-T: 0.844 ± 0.054, MAE-LP: 0.539 ± 0.053, MAE-FT: 0.798 ± 0.042, SimCLR-LP: 0.904 ± 0.008, and SimCLR-FT: 0.929 ± 0.028 (Kaczmarek et al., 12 Sep 2025). The paper explicitly states that this is about a 6% AUC improvement over the next-best model. For sex classification on IXI, the corresponding AUC values are 0.988 ± 0.005, 0.876 ± 0.014, 0.839 ± 0.026, 0.864 ± 0.025, 0.896 ± 0.006, and 0.991 ± 0.004, respectively. For age regression on IXI, the MAE values are 4.85 ± 0.13, 7.53 ± 0.25, 15.28 ± 0.18, 8.53 ± 0.28, 5.90 ± 0.16, and 4.35 ± 0.22, respectively.

Several relative conclusions are reported directly in the paper. SimCLR fine-tuning beats supervised training from scratch even though the supervised models are trained for 300 epochs and SimCLR fine-tuning uses only 100 epochs. SimCLR beats MAE consistently, suggesting that for this dataset size and MRI modality, contrastive learning with a CNN backbone is more effective than a ViT-based masked reconstruction objective. The ViT-based models struggle more overall, which the authors attribute to limited data and greater hyperparameter sensitivity of ViTs (Kaczmarek et al., 12 Sep 2025).

The low-label analysis uses 20%, 40%, 60%, 80%, and 100% of the labeled downstream data. Performance usually drops only modestly until the training set reaches 20%. For Alzheimer’s classification, even with just 20% of the labeled data, corresponding to only about 10 samples per class, the SimCLR model still outperforms all supervised baselines, and the AUC only drops by about 4% as the training data shrinks despite class imbalance in validation and test (Kaczmarek et al., 12 Sep 2025). This suggests that the learned representation remains effective in the low-label regime that motivates foundation-model approaches in clinical imaging.

7. Accessibility, limitations, and relation to prior 3D SSL

The work emphasizes public accessibility. Code is publicly available on GitHub, the trained SimCLR foundation model is released publicly, preprocessing uses public tools and a reproducible pipeline, and the stated repository is https://github.com/emilykaczmarek/3D-Neuro-SimCLR (Kaczmarek et al., 12 Sep 2025). The reliance on public datasets and public preprocessing software is presented as a reproducibility advantage relative to prior work.

The paper also implies several limitations. The pretraining set is large for medical imaging, but still relatively small compared with natural-image foundation models, which likely explains why ViTs and MAEs do not shine in this setting. Pretraining is restricted to T1 structural MRI only, so the work does not address multimodal MRI. Downstream evaluation covers only four tasks, although they are diverse. The MAE comparison is implemented carefully, but the authors suggest that ViT/MAE may need more optimization and/or more data. Because the work focuses on publicly available datasets, some clinically important private cohorts are not included (Kaczmarek et al., 12 Sep 2025).

In the broader 3D self-supervised literature, 3D-Neuro-SimCLR can be situated as a foundation-model-scale extension of earlier volumetric SimCLR approaches. Earlier work showed that 3D SimCLR can be used for medical segmentation by pretraining a 3D encoder on unlabeled volumes, discarding the projection head, attaching a decoder, and fine-tuning on labeled data; that line of work also reported gains in annotation-starved settings and explored Monte Carlo Dropout as an uncertainty-aware inference mechanism (Ali et al., 2021). Relative to that predecessor, 3D-Neuro-SimCLR broadens the objective from task-specific segmentation to a public, high-resolution, multi-disease brain MRI representation model intended for transfer across both disease and healthy-cohort tasks.