SSL-AD: Self-Supervised Alzheimer’s MRI

• SSL-AD is a self-supervised framework that learns spatiotemporal 3D MRI representations to address Alzheimer’s prediction challenges including variable scan lengths and irregular intervals.
• It combines temporal order prediction with a contrastive learning module to enhance spatial feature extraction and ensure adaptability across multi-scan inputs.
• Empirical evaluations show SSL-AD outperforms supervised baselines in diagnosis classification, conversion detection, and future conversion prediction across multiple datasets.

Searching arXiv for the term and closely related papers to ground the article and clarify acronym usage. arXiv search query: "SSL-AD" SSL-AD most specifically denotes “Spatiotemporal Self-Supervised Learning for Generalizability and Adaptability Across Alzheimer’s Prediction Tasks and Datasets,” a self-supervised framework for longitudinal 3D brain MRI that learns reusable spatiotemporal representations from unlabeled data and then fine-tunes the same backbone for diagnosis classification, conversion detection, and future conversion prediction. In the literature, the acronym is not unique: it also appears in an unrelated secure-systems context through Stacco’s differential side-channel analysis of SSL/TLS vulnerabilities in enclaves, and it is used more loosely as a shorthand for self-supervision in anomaly detection. In current Alzheimer’s prediction research, however, SSL-AD refers to the MRI-based spatiotemporal SSL framework introduced for cross-task, cross-dataset, and variable-length longitudinal modeling (Kaczmarek et al., 12 Sep 2025, Xiao et al., 2017, Akoglu et al., 2023).

1. Scope, nomenclature, and position in the literature

SSL-AD is situated in the recent line of self-supervised representation learning for neurodegenerative disease modeling. Its stated objective is to learn general spatiotemporal MRI representations via self-supervision on large unlabeled data, so that one pre-trained model can be fine-tuned for multiple downstream tasks while remaining adaptable to any number of input scans from 1 to 4 and to irregular time intervals without retraining the backbone (Kaczmarek et al., 12 Sep 2025).

A recurring source of confusion is acronym collision. “SSL-AD” in Stacco refers to a differential side-channel analysis workflow for detecting SSL/TLS oracle vulnerabilities in SGX enclaves, not to Alzheimer’s disease modeling (Xiao et al., 2017). Separately, a vision paper on anomaly detection uses “SSL-AD” as a generic abbreviation for self-supervised learning for anomaly detection, again unrelated to MRI-based disease prediction (Akoglu et al., 2023). In Alzheimer’s research proper, a nearby but distinct method is TE-SSL, which incorporates event and time-to-event signals into self-supervised pretraining for downstream survival analysis rather than the multi-task longitudinal prediction setup used by SSL-AD (Thrasher et al., 2024).

This terminological ambiguity suggests that the most precise usage is to reserve “SSL-AD” for the 2025 Alzheimer’s paper when discussing spatiotemporal MRI representation learning, and otherwise to expand the acronym explicitly.

2. Problem formulation, data regime, and preprocessing

SSL-AD is motivated by three stated challenges in Alzheimer’s prediction: very limited labeled longitudinal MRI data, dataset shifts arising from differences in MRI protocols across cohorts, and variability in both the number of visits and the intervals between them. Conventional deep learning methods are described as label-intensive, task-specific, and rigid with respect to input length (Kaczmarek et al., 12 Sep 2025).

The framework’s three explicit objectives are to learn general spatiotemporal MRI representations via self-supervision on large unlabeled data, to allow a single pre-trained model to be fine-tuned for diagnosis classification, conversion detection, and future conversion prediction, and to ensure adaptability to any number of input scans from 1 to 4 with irregular time intervals. The abstract reports four publicly available datasets comprising 3,161 patients for pre-training, while the detailed summary identifies HABS-HD, MCSA, NIFD, and an ADNI pre-training split. This suggests that the paper distinguishes the aggregate pre-training design from a specific subject-count convention used in the detailed summary (Kaczmarek et al., 12 Sep 2025).

For downstream evaluation, the held-out ADNI split is reported as 2,074 training patients, 471 validation patients, and 483 test patients. Standardized 3D MRI preprocessing is performed with TurboPrep: N4 bias–field correction, skull-stripping with SynthStrip, linear and diffeomorphic registration to a common template, tissue segmentation with SynthSeg, intensity normalization, and resampling to $1\times1\times1\ \mathrm{mm}^3$, yielding $[193\times229\times193]$, followed by depth-first permutation and cropping to $[150\times192\times192]$ (Kaczmarek et al., 12 Sep 2025).

The SSL augmentations are implemented in MONAI. Affine perturbations with rotation $\pm 0.34$ rad, translation $\pm 15$ voxels, and scale $0.9$–$1.3$ are applied consistently across all timepoints, while Gaussian smoothing with $\sigma \in [0.25,1.5]$ and Gaussian noise with $\sigma \in [0.05,0.09]$ are applied independently. Inputs are normalized by the overall training-set standard deviation. Downstream fine-tuning uses stronger spatial transforms (Kaczmarek et al., 12 Sep 2025).

3. Architecture and temporal design

The backbone encoder $f_\theta(\cdot)$ is a 3D ResNet-18 whose penultimate feature is 512-dimensional. For each MRI volume $[193\times229\times193]$0, the spatial encoder produces

$[193\times229\times193]$1

These per-timepoint features are concatenated as

$[193\times229\times193]$2

and passed to a 2-layer MLP $[193\times229\times193]$3 that serves as the temporal module (Kaczmarek et al., 12 Sep 2025).

A central design constraint is variable-length input handling. SSL-TOV, the order-verification variant, fixes $[193\times229\times193]$4, pads shorter sequences with all-zero volumes, and uses a single binary head. SSL-TOP and SSL-TOPC, the order-prediction variants with and without a contrastive term, avoid padding and instead maintain three separate classifier heads $[193\times229\times193]$5, $[193\times229\times193]$6, and $[193\times229\times193]$7 for $[193\times229\times193]$8; at each epoch, only one sequence length is sampled (Kaczmarek et al., 12 Sep 2025).

Irregular time intervals are incorporated during downstream fine-tuning for multi-scan inputs by concatenating the scalar differences $[193\times229\times193]$9 to the temporal MLP input. This design is intended to preserve backbone reuse while allowing temporal reasoning over sequences with nonuniform spacing (Kaczmarek et al., 12 Sep 2025).

The architectural choice differentiates SSL-AD from TE-SSL. TE-SSL uses a 3D CNN encoder and a projection head for contrastive pretraining, then replaces the projection head with a DeepHit survival head for time-to-event analysis; SSL-AD instead retains a task-general spatiotemporal backbone intended for classification, conversion detection, and prognosis across different input cardinalities (Thrasher et al., 2024).

4. Self-supervised objectives and optimization

SSL-AD combines temporal order prediction with single-timepoint contrastive learning. In temporal order prediction, a sequence $[150\times192\times192]$0 of $[150\times192\times192]$1 augmented scans is permuted and assigned a one-hot label over the $[150\times192\times192]$2 possible permutations. The classification head $[150\times192\times192]$3 outputs $[150\times192\times192]$4, and the loss is

$[150\times192\times192]$5

where $[150\times192\times192]$6 is the ground-truth one-hot component and $[150\times192\times192]$7 the predicted probability (Kaczmarek et al., 12 Sep 2025).

The contrastive component uses only the first timepoint in the sequence. Two independently augmented views $[150\times192\times192]$8 and $[150\times192\times192]$9 are created, and the projected embeddings are

$\pm 0.34$0

With cosine similarity

$\pm 0.34$1

the NT-Xent loss for a positive pair in a batch of $\pm 0.34$2 sequences is

$\pm 0.34$3

and the symmetric contrastive objective is

$\pm 0.34$4

The total SSL pre-training loss is

$\pm 0.34$5

Equal weighting $\pm 0.34$6 is used by default (Kaczmarek et al., 12 Sep 2025).

Pre-training samples all pairs, triplets, and quadruplets randomly each epoch from 3,161 subjects, with approximately $\pm 0.34$7 sampling across sequence lengths. The batch size is 16 sequences; optimization uses Adam with $\pm 0.34$8, $\pm 0.34$9, weight decay $\pm 15$0, initial learning rate $\pm 15$1 decayed by $\pm 15$2 every 50 epochs, and 200 total epochs with early stopping on held-out aggregate validation. For downstream tasks, the final MLP head is replaced and reinitialized; in single-MRI tasks, $\pm 15$3 is frozen for the first 5 epochs and then unfrozen. Fine-tuning uses learning rate $\pm 15$4, batch size 8, and 50 epochs with early stopping (Kaczmarek et al., 12 Sep 2025).

5. Downstream tasks and empirical performance

The downstream task suite contains three families. Diagnosis classification predicts CN versus MCI versus AD using 1, 2, or 3 scans separated by 1 to 2.5 years. Conversion detection uses 2 scans to classify stable versus converted for CN$\pm 15$5MCI or MCI$\pm 15$6AD. Future conversion prediction uses a single baseline scan to predict whether conversion occurs within the next 1 to 2.5 years for CN$\pm 15$7MCI or MCI$\pm 15$8AD. The primary evaluation metric is area under the ROC curve (AUC), with accuracy, sensitivity, and specificity also available on demand (Kaczmarek et al., 12 Sep 2025).

The study compares a fully supervised ResNet-18 baseline against SSL-TOV, SSL-TOP, and SSL-TOPC. The detailed AUCs show that SSL-TOPC outperforms supervised learning on six of seven downstream tasks, while SSL-TOP attains the best score on 3-scan classification and MCI$\pm 15$9AD conversion detection, and SSL-TOV attains the best score on MCI$0.9$0AD prediction (Kaczmarek et al., 12 Sep 2025).

Task	Supervised ResNet-18 AUC	Best SSL result
Classification (1 MRI)	$0.9$1	SSL-TOPC: $0.9$2
Classification (2 MRIs)	$0.9$3	SSL-TOPC: $0.9$4
Classification (3 MRIs)	$0.9$5	SSL-TOP: $0.9$6
CN$0.9$7MCI Conversion (2 MRIs)	$0.9$8	SSL-TOPC: $0.9$9
MCI$1.3$0AD Conversion (2 MRIs)	$1.3$1	SSL-TOP: $1.3$2
CN$1.3$3MCI Prediction (1 MRI)	$1.3$4	SSL-TOPC: $1.3$5
MCI$1.3$6AD Prediction (1 MRI)	$1.3$7	SSL-TOV: $1.3$8

The pattern of results is technically informative. The contrastive term appears especially helpful in single-MRI settings, whereas explicit permutation prediction appears especially helpful when richer temporal context is available. This interpretation is directly supported by the ablation results rather than inferred solely from headline AUCs (Kaczmarek et al., 12 Sep 2025).

6. Generalization, ablations, limitations, and related directions

The generalization and adaptability claims are explicit. Cross-dataset evaluation pre-trains on aggregated HABS-HD, MCSA, NIFD, and ADNI-train and tests on held-out ADNI, which is reported to demonstrate robustness to cohort shift. Performance increases monotonically with the number of scans for all models; for example, the supervised baseline rises from $1.3$9 to $\sigma \in [0.25,1.5]$0 to $\sigma \in [0.25,1.5]$1, and SSL-TOPC rises from $\sigma \in [0.25,1.5]$2 to $\sigma \in [0.25,1.5]$3 to $\sigma \in [0.25,1.5]$4. By sampling 1 to 2.5 year gaps during SSL, the fine-tuned models remain stable across 12- to 30-month separations (Kaczmarek et al., 12 Sep 2025).

The ablations refine the functional role of each pretext component. Moving from binary order verification in SSL-TOV to full permutation prediction in SSL-TOP yields a substantial gain on 3-scan classification, from $\sigma \in [0.25,1.5]$5 to $\sigma \in [0.25,1.5]$6, with smaller changes on 1- and 2-scan tasks. Adding the contrastive term in SSL-TOPC boosts 1-MRI classification from $\sigma \in [0.25,1.5]$7 to $\sigma \in [0.25,1.5]$8 and CN$\sigma \in [0.25,1.5]$9MCI prediction from $\sigma \in [0.05,0.09]$0 to $\sigma \in [0.05,0.09]$1, indicating that the contrastive objective chiefly improves spatial feature learning when temporal context is absent. Early tests with $\sigma \in [0.05,0.09]$2 showed slightly diminished AUC for multi-scan tasks when $\sigma \in [0.05,0.09]$3, suggesting that the temporal objective must remain primary in multi-scan settings (Kaczmarek et al., 12 Sep 2025).

The stated limitations are also specific. Pre-training is capped at 4 scans, so extending to longer sequences requires new permutation strategies. There is no external validation on non-ADNI test cohorts. Potential future work includes integrating PET and CSF biomarkers into the SSL framework (Kaczmarek et al., 12 Sep 2025).

Within Alzheimer’s disease progression analysis, SSL-AD and TE-SSL illustrate two different strategies for enriching self-supervised representations. SSL-AD emphasizes temporal order prediction, single-timepoint contrastive learning, and flexible fine-tuning across seven downstream tasks, whereas TE-SSL injects event and time-to-event labels directly into the SSL loss and reports the best downstream survival-analysis performance at $\sigma \in [0.05,0.09]$4 and $\sigma \in [0.05,0.09]$5 for the configuration $\sigma \in [0.05,0.09]$6 (Thrasher et al., 2024). A plausible implication is that the choice between these frameworks depends on whether the target application is multi-task longitudinal prediction or explicitly survival-oriented modeling.