MCL-AD: Multi-Level Connectome in AD

• MCL-AD is a contextual label used to denote methods for learning brain connectivity for Alzheimer’s diagnosis, MCI stratification, and conversion prediction.
• It involves multi-level generated connectomes employing spatio-temporal feature extraction and graph convolution networks to analyze rs-fMRI and MRI data.
• The term’s ambiguity—with application in AD research and zero-shot 3D anomaly detection—highlights the need for clear contextual disambiguation.

MCL-AD is not a single stable term across the arXiv literature. In the supplied corpus, it denotes several distinct research programs, most of them centered on Alzheimer’s disease (AD) and mild cognitive impairment (MCI), but one of them unrelated to neuroimaging and instead concerned with zero-shot 3D anomaly detection. In the AD literature represented here, the label is used descriptively for methods that learn functional or structural connectivity representations for AD diagnosis, MCI stratification, or MCI-to-AD conversion prediction, with the most explicit formulation being a multi-level generated connectome pipeline instantiated by MLC-GCN on resting-state fMRI (Zhu et al., 2024). This suggests that MCL-AD is best treated as a contextual label rather than a canonical author-defined acronym.

1. Nomenclature and scope

The supplied papers attach closely related strings to different technical objects. Some are AD-focused and some are not.

Usage in the supplied corpus	Core task	Representative paper
Multi-level connectome learning for AD	rs-fMRI classification of NC, MCI subgroups, and AD with multi-level generated connectomes and GCNs	(Zhu et al., 2024)
Multi-class clinical group learning for AD	Representative functional connectivity learning across NC, SMC, sMCI, pMCI, and AD	(Zhang et al., 2021)
MRI-derived MCI-to-AD conversion learning	Baseline MRI prognosis using surrogate conversion labels derived from an AD-Control classifier trajectory	(Aksu et al., 2011)
Longitudinal structural-network conversion prediction	T1-derived structural brain networks, healthy-aging trajectory modeling, and residual-based prediction of MCI-to-AD conversion	(Wei et al., 2022)
Multimodal collaboration learning for anomaly detection	Zero-shot 3D anomaly detection from point clouds, RGB images, and text priors	(Li et al., 12 Sep 2025)

Within the AD-focused subset, the common thread is not a shared acronym but a shared problem family: learning clinically informative latent structure from brain connectivity or brain morphology to distinguish normal controls from pathological or prodromal states. By contrast, the exact-title paper “MCL-AD: Multimodal Collaboration Learning for Zero-Shot 3D Anomaly Detection” is semantically unrelated to Alzheimer’s disease despite the identical string (Li et al., 12 Sep 2025).

A frequent misconception is therefore terminological rather than methodological: the same label can refer either to connectome-based AD modeling or to a later anomaly-detection framework. The supplied evidence does not support treating these as one lineage.

2. Multi-level connectome learning for Alzheimer’s disease

In the supplied material, the most fully specified AD usage is the multi-level generated connectome framework implemented as MLC-GCN, a stack of spatio-temporal feature extraction and graph generation followed by per-level GCN encoding for AD analysis from resting-state fMRI (Zhu et al., 2024). The clinical task is multi-class classification. On ADNI, the task is four-class with NC, EMCI, LMCI, and AD. On OASIS-3, it is three-class with NC, MCI, and AD.

The motivation is that AD-related functional connectivity alterations occur across multiple spatial and temporal scales. Traditional single-level FC matrices or shallow machine learning may miss hierarchical spatio-temporal structure, while population-level GCNs may face scalability and generalization issues. MLC-GCN addresses this by generating subject-specific connectomes at multiple spatio-temporal levels and learning a separate GCN encoder at each level before fusing the resulting graph embeddings (Zhu et al., 2024).

The data and acquisition settings are explicitly heterogeneous. ADNI contributes 643 rs-fMRI scans acquired on a 3T Philips scanner, with NC $=187$, EMCI $=182$, LMCI $=156$, and AD $=118$. OASIS-3 contributes 900 rs-fMRI scans acquired on a 3T Siemens scanner, with NC $=570$, MCI $=225$, and AD $=105$. The model is validated independently on both cohorts with 5-fold stratified cross-validation, and the supplied summary emphasizes different scanners and acquisition protocols as part of the robustness claim (Zhu et al., 2024).

Cohort	Labels	Acquisition details
ADNI	NC, EMCI, LMCI, AD	3T Philips; TR/TE $=3000/30$ ms; matrix $64\times64$; voxel size $3.3\times3.3\times3.3\ \mathrm{mm}^3$; 48 slices
OASIS-3	NC, MCI, AD	3T Siemens; TR/TE $=182$0 ms; matrix $=182$1; voxel size $=182$2; 36 slices

Preprocessing is performed with the BRANT toolkit and includes slice timing correction, realignment to the first volume, spatial normalization to MNI space, nuisance regression, temporal band-pass filtering at $=182$3, and parcellation with the Brainnetome Atlas into 273 ROIs. ROI time series are defined as the average BOLD signal across all voxels in each ROI at each time point. The framework uses static FC rather than dynamic FC, and scrubbing is not described. The lowest-level connectome is the subject-specific Pearson correlation matrix $=182$4 computed from the ROI time series, with $=182$5 (Zhu et al., 2024).

3. Architectural formulation of the multi-level generated connectome

The MLC-GCN pipeline has two major blocks: a multi-graph generation block and a GCN prediction block (Zhu et al., 2024). For a subject with ROI time series $=182$6, an input embedding layer applies a 1D-CNN over time, followed by a linear layer and sinusoidal positional encoding, to produce $=182$7. The supplied formulation is

$=182$8

The stack of spatio-temporal feature extraction modules is recursive. At level $=182$9, the spatial feature extraction path uses a Transformer encoder on $=156$0, producing $=156$1. The temporal feature extraction path uses a DLinear-style decomposition with

$=156$2

followed by

$=156$3

Fusion gives

$=156$4

with $=156$5 and $=156$6 for $=156$7 (Zhu et al., 2024).

The connectome hierarchy consists of one lowest-level Pearson FC graph and $=156$8 learned higher-level adjacency matrices. The lowest level is the standard FC matrix

$=156$9

forming $=118$0. At higher levels,

$=118$1

These graphs are symmetric, weighted, and undirected. No thresholding or sparsification is applied during training, although the top 1% strongest edges are shown for visualization. For propagation, self-loops are added: $=118$2 The supplied summary states that degree normalization is not described and that the implementation uses $=118$3 directly (Zhu et al., 2024).

The GCN prediction block trains a separate two-layer GCN for each level $=118$4, with no weight sharing across levels. Node features are initialized identically for all levels as the rows of the Pearson FC matrix,

$=118$5

Propagation is

$=118$6

After the second graph-convolution layer, an MLP produces a per-level graph embedding $=118$7. Multi-level fusion is simple concatenation,

$=118$8

and final prediction uses

$=118$9

Training combines cross-entropy with an intra-group regularizer that encourages per-class adjacency consistency at each learned level. The total loss is

$=570$0

with $=570$1. Optimization uses AdamW with learning rate $=570$2, weight decay $=570$3, dropout $=570$4, up to 300 epochs, 5-fold stratified cross-validation, and Mixup augmentation. The STFE depth was tested at $=570$5, and the implementation is in PyTorch on a single NVIDIA RTX 4090 GPU (Zhu et al., 2024).

4. Validation, ablation, and interpretability

The reported evaluation metrics are Accuracy, AUC, specificity, sensitivity, and F1-score, given as mean $=570$6 standard deviation over 5 folds. Statistical tests and confidence intervals are not reported (Zhu et al., 2024).

For binary NC versus AD classification on ADNI, the best reported variant, MLC-GCN$=570$7, achieves Acc $=570$8, AUC $=570$9, Spe $=225$0, Sen $=225$1, and F1 $=225$2. For four-class ADNI classification, MLC-GCN$=225$3 achieves Acc $=225$4, AUC $=225$5, Spe $=225$6, Sen $=225$7, and F1 $=225$8, which the supplied summary describes as approximately 6 percentage points absolute improvement over LG-GNN with Acc $=225$9. For three-class OASIS-3 classification, MLC-GCN$=105$0 achieves Acc $=105$1, AUC $=105$2, Spe $=105$3, Sen $=105$4, and F1 $=105$5 (Zhu et al., 2024).

Setting	Reported result for MLC-GCN$=105$6	Comparison stated in the supplied summary
ADNI, NC vs AD	Acc $=105$7, AUC $=105$8	Outperforms a wide range of RF, SVM, CNN/DNN, GCN, generated-graph GNN, and hybrid baselines
ADNI, 4-class	Acc $=105$9, AUC $=3000/30$0	About 6 percentage points absolute improvement over LG-GNN
OASIS-3, 3-class	Acc $=3000/30$1, AUC $=3000/30$2	Matches or exceeds strong baselines such as LG-GNN

The ablation results are structurally informative. Removing either the spatial feature extraction path or the temporal feature extraction path reduces performance, with the larger drop observed when spatial modeling is removed. Using only the last-level graph, or replacing the multi-level hierarchy with repeated Pearson graphs, also degrades performance. The embedding length $=3000/30$3 matters: the reported summary states that performance peaks around $=3000/30$4, that $=3000/30$5 is a good range, that too small a value loses information, and that too large a value adds noise (Zhu et al., 2024).

Interpretability is derived from average generated connectomes and node importance computed as the sum of incident edge weights in the average graph. Higher-level generated graphs are reported as sparser than raw Pearson FC and as concentrating on prefrontal and temporal cortices. Across ADNI and OASIS-3, the top 20 discriminative ROIs consistently include the superior frontal gyrus, middle frontal gyrus, inferior frontal gyrus, paracentral lobule, superior temporal gyrus, and middle temporal gyrus. Visualization with BrainNet Viewer of the top 1% strongest edges shows hub-like patterns in these regions (Zhu et al., 2024). A plausible implication is that the hierarchical graph construction is not merely improving classification but also filtering the raw FC into subnetworks that are more clinically aligned.

5. Other AD-related frameworks described with the same shorthand

A separate line of work uses a deep encoder–decoder with class prototypes to learn representative functional connectivity patterns across multiple AD-related groups. In “Representative Functional Connectivity Learning for Multiple Clinical groups in Alzheimer’s Disease” the model maps vectorized FC matrices into a latent space, learns a trainable prototype $=3000/30$6 for each class $=3000/30$7, and reconstructs representative FCs by decoding these prototypes back into FC space as $=3000/30$8. The paper does not introduce a formal acronym, and the supplied summary explicitly notes that “MCL-AD” is a shorthand rather than an author-provided name. The model achieves approximately $=3000/30$9 multi-class accuracy, with $64\times64$0 and $64\times64$1, and identifies frontal- and temporal-lobe edges as among the most variable across groups. It also reports that sMCI exhibits more positive correlation values than pMCI (Zhang et al., 2021).

A second formulation concerns prognosis rather than diagnosis. “An MRI-Derived Definition of MCI-to-AD Conversion for Long-Term, Automatic Prognosis of MCI Patients” defines conversion by whether follow-up MRIs from an MCI subject cross the AD side of a linear AD-Control SVM margin, either directly or by a fitted trajectory $64\times64$2. The resulting converter-by-trajectory versus nonconverter-by-trajectory labels are then used to train a baseline-MRI prognostic SVM. Reported prognostic accuracy for the trajectory-based definition ranges from $64\times64$3 to $64\times64$4, whereas the supplied summary reports $64\times64$5 to $64\times64$6 for CDR-based conversion definitions in comparable settings. The same summary also emphasizes that there are no definitive labeled examples of MCI conversion in ADNI and explicitly discusses the risk of circularity, since surrogate labels are derived from an AD-Control imaging classifier (Aksu et al., 2011).

A third framework predicts MCI-to-AD conversion at an 18-month horizon by generating structural brain networks from T1 MRI under diffusion-MRI guidance, then modeling deviation from a learned healthy aging trajectory. In this pipeline, T1 and FA features are aligned with a contrastive loss, graph embeddings are produced with three GATConv layers and global pooling, a VAE-style model predicts healthy-aging graph features, and an LSTM predicts future deviation vectors that are mapped to a conversion probability. The reported performance for conversion prediction is accuracy $64\times64$7, sensitivity $64\times64$8, and specificity $64\times64$9. The method requires diffusion MRI during training but uses T1 MRI alone at inference (Wei et al., 2022).

Taken together, these works show that the AD-related use of MCL-AD spans at least three distinct targets: group-level FC representation learning, surrogate-label prognostication from structural MRI trajectories, and longitudinal deviation modeling from T1-derived structural networks. They are methodologically adjacent but not architecturally interchangeable.

6. Ambiguity beyond Alzheimer’s disease, limitations, and likely directions

The broader supplied corpus makes the naming problem explicit. Outside Alzheimer’s disease, “MCL” is used for Monte Carlo Localization (Chen et al., 2020, Eisoldt et al., 2023), Mobile Collaborative Learning [(Razaque et al., 2012); (Razaque et al., 2012)], and Multinoulli Contrastive Loss in product quantization (Xiao et al., 2021). The exact string “MCL-AD” is also used by a later paper on zero-shot 3D anomaly detection, where it expands to “Multimodal Collaboration Learning for Zero-Shot 3D Anomaly Detection” and denotes a framework based on point clouds, RGB images, and text prompts rather than neuroimaging (Li et al., 12 Sep 2025). This suggests that any use of the term in citation networks or bibliographic tools requires immediate local disambiguation.

Within the connectome-centric AD formulation, several limitations are explicitly noted. Class imbalance exists and is not handled with imbalance-aware losses. Nuisance-regression specifics and scrubbing are not detailed. The model uses static FC rather than dynamic FC, and the implementation uses $3.3\times3.3\times3.3\ \mathrm{mm}^3$0 without degree normalization. Explainability is based on edge weights in averaged graphs rather than perturbation- or gradient-based attribution, and no formal domain adaptation is used despite cross-cohort validation (Zhu et al., 2024). The representative-FC prototype model does not explicitly constrain decoder outputs to be symmetric or to preserve unit diagonals, relying instead on a structural regularizer (Zhang et al., 2021). The MRI-derived conversion framework depends on surrogate labels and therefore inherits label-noise and circularity concerns (Aksu et al., 2011). The T1-based longitudinal structural-network framework does not report explicit harmonization across ADNI sites and does not address irregular sampling or missing visits (Wei et al., 2022).

The future directions stated across these papers are convergent. For multi-level connectome learning, the supplied summary names multi-modal integration, longitudinal prediction, dynamic structure learning, stronger GNN encoders, and applications to other diseases or clinical outcomes (Zhu et al., 2024). For representative FC learning, larger and harmonized multi-site cohorts and formal statistical testing of edges are proposed (Zhang et al., 2021). For MRI-derived conversion, the recommendations include multimodal biomarkers, survival-style modeling, calibration, and external validation (Aksu et al., 2011). For longitudinal structural-network prediction, the existing design already points toward a broader program in which healthy-aging trajectories and subject-specific deviations become central latent variables for clinical forecasting (Wei et al., 2022).

In that sense, the most coherent AD-specific reading of MCL-AD is not as a fixed acronym but as a family resemblance among methods that learn clinically structured latent representations of brain connectivity or morphology, often with explicit graph, prototype, or trajectory mechanisms, for diagnosis and prognosis in the AD continuum.