Anatomical Grounding and Leakage-Aware Multimodal Contrastive Learning for Alzheimer's Disease Classification from Structural MRI
Paul-Gabriel Nicolae, Irina Georgiana Mocanu
cs.CV · cs.AI
Quick verdict
A careful leakage-aware study showing that aligning an MRI encoder to regional volumes rather than cognitive scores yields a far better image-only representation for MCI vs CN, but the 21-point headline rests on 9 test scans from a single run on a 63-scan test set.
Summary
This paper audits two recurring problems in MRI-based Alzheimer's staging using a deliberately lightweight 2D pipeline on 1,075 baseline T1-weighted ADNI-1 scans (982 subjects; roughly 289 CN / 507 MCI / 279 AD), with subject-level splits and a balanced test set of 21 scans per class (63 for three-way, 42 per binary task). The encoder is an ImageNet-pretrained ResNet18 applied per axial slice (indices 62-152, 90 slices), with only layer4 fine-tuned, aggregated by a single 8-head Transformer layer into a 128-d embedding. First, FastSurfer segmentations are used to build an anatomical reference: YOLOv8 detectors and instance-segmentation models trained on 21 and 20 AD-relevant structures respectively reach mAP50 of 0.951-0.969, and a qualitative Grad-CAM comparison shows the image classifier attends both to meaningful tissue (lateral ventricles, temporal lobe) and, on many slices, to the skull, orbits, neck and background. Second, a CLIP-style image-tabular contrastive framework is adapted with a TabTransformer-style tabular branch, and ADNIMERGE variables are organized along a label-leakage spectrum: cognitive scores (CDR-SB, ADAS11, MMSE, RAVLT-immediate) are part of the diagnostic criteria, while regional volumes (hippocampus, whole brain, entorhinal, middle temporal) are only atrophy correlates. Fusion with cognitive scores gives 87.3% three-way accuracy, which the authors correctly treat as a leakage-driven upper bound; fusion with volumes gives 73.0%. The main finding concerns the representation rather than the fused number: on MCI vs CN, the image-only head reaches 73.8% when the encoder is contrastively aligned to volumes but only 52.4% when aligned to cognitive scores, with no tabular input at inference. Restricting input to a per-subject medial temporal lobe crop raises image-only three-way accuracy from 58.7% to 65.1%. Limitations are stated plainly and are severe: 63 test scans, single runs, and a 21-point gap equivalent to nine scans at p about 0.04 unpaired.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Deep MRI classifiers for AD staging frequently attend to anatomically irrelevant regions (skull, orbits, neck, background), as shown by comparing Grad-CAM maps against YOLOv8 detections of AD-relevant structures.
- Multimodal performance obtained by fusing cognitive scores (87.3% three-way) largely reflects label leakage, because MMSE and CDR-SB are part of the ADNI diagnostic criteria, and should be reported as an upper bound rather than an imaging result.
- The contrastive target determines what the image encoder learns: on MCI vs CN the image-only head reaches 73.8% when aligned to regional volumes and 52.4% when aligned to cognitive scores, with no tabular input at inference.
- Restricting the input to a per-subject medial temporal lobe crop raises image-only three-way accuracy from 58.7% to 65.1% and MCI vs CN accuracy from 73.8% to 81.0% without changing the encoder.
- Fused accuracy alone cannot distinguish improved use of imaging from access to the labelling rule, so the image-only head should always be reported alongside the fused one.
Strengths
Intellectually honest framing of the leakage problem, with the strongest number explicitly demoted
Section 6.3 argues from the ADNI-1 diagnostic criteria (Petersen et al. 2010) that MMSE and CDR-SB are part of the rule that produced the label, and Table 3 consequently reports the 87.3% cog-fusion result as an 'upper bound, leakage' rather than a result. Few papers in this literature do this; most report the leaky number as state of the art.
The main finding is a clean controlled comparison on the representation, not on a leaderboard number
Section 8 varies exactly one thing — which tabular group the CLIP loss aligns the image embedding to — holding architecture, training data, subject split, two-phase schedule, hyperparameters and test set fixed. The image-only head takes no tabular input at inference, so 73.8% versus 52.4% on MCI vs CN isolates an effect on the visual representation. The inversion against fused accuracy (87.3% cog vs 73.0% vol three-way) is the sharpest point in the paper: the configuration that looks better end to end learned the weaker image encoder.
Careful statistical reporting given the constraints, including Wilson intervals on every result
Anatomical priors used twice and coherently: once to audit, once to fix
Explicit refusal to claim superiority over published work, with the protocol differences enumerated
Weaknesses
A 63-scan test set and single runs make nearly every comparison underpowered
The MTL crop and the vol contrastive target derive from the same structures, so the two interventions are not independent
The 2D slice encoder was chosen knowing it is about 20 points weaker than the 3D alternative
Leakage is argued from the diagnostic criteria but never measured directly
Only two of five modality groups are evaluated, and one planned comparison is missing
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper takes on two problems that the ADNI classification literature routinely glosses over, and it does so with unusual candour. Problem one: a CNN fed whole axial slices has no reason to stay inside the head, and Grad-CAM evidence here shows attention landing on skull, orbits, neck and background as well as on ventricles and temporal lobe. Problem two: ADNI ships rich clinical tables, and the columns that produce the biggest accuracy gains — MMSE, CDR-SB — are the same instruments used to assign the diagnostic label in the first place. Five claims follow: FastSurfer-derived detections give a usable anatomical reference (Table 2, mAP50 0.951-0.969); cog fusion at 87.3% is a leakage upper bound, not an imaging result; the contrastive target determines what the image encoder learns; an MTL crop improves image-only accuracy; and fused numbers should never be reported without the image-only head.
This is a modest, well-scoped study and the authors say so. They are explicit that the 3D backbone was about 20 points better in a preliminary comparison and that the 2D design was kept for interpretability reasons. That choice is defensible for the questions asked, but it does mean the absolute accuracies sit below published 3D work and should be read as a vehicle for the audits rather than as a competitive system.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is label leakage in Alzheimer's disease prediction from ADNI data?
ADNI assigns the CN, MCI and AD labels using clinical criteria that include the MMSE score, the global Clinical Dementia Rating and memory testing. Several columns in the ADNIMERGE table, such as MMSE and the CDR sum of boxes, are therefore part of the rule that produced the label. A model given those columns can score highly by re-deriving the rule rather than by extracting information from the MRI scan. This paper treats accuracy obtained with cognitive scores as an upper bound and a sanity check, not as evidence about imaging.
Why does aligning the image encoder to brain volumes work better than aligning it to cognitive scores?
The authors' interpretation is that the contrastive target determines what the image branch must encode in order to match it. Regional volumes such as hippocampal and entorhinal size live in a spatial, structural feature space, so the image embedding can only be pulled toward them by representing atrophy. Cognitive scores offer no spatial structure, and because they nearly determine the label, the fusion head can solve the task from the tabular branch alone, leaving the image encoder with no useful gradient.
How accurate is the model?
On a balanced 63-scan test set, the image-only model on full slices reaches 58.7% three-way accuracy and 85.7% on AD versus CN. Fusing regional volumes gives 73.0% three-way; fusing cognitive scores gives 87.3%, which the authors treat as leakage-driven. Restricting the input to a medial temporal lobe crop raises image-only three-way accuracy to 65.1%. The authors stress these are single runs on a small test set and are not comparable to published numbers, which use different protocols.
Does the model look at the right parts of the brain?
Partly. Comparing Grad-CAM maps with YOLOv8 detections of anatomically defined structures, the authors find that the classifier does attend to the lateral ventricles and, on lower slices, to temporal lobe structures including the hippocampus. But on a substantial number of slices attention falls on the skull, orbits, neck or background, so correct predictions can be supported by anatomically irrelevant evidence. That observation motivated restricting the input to a medial temporal lobe crop.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 (architecture diagram, Grad-CAM overlays, the MCI/CN bar chart and the MTL crop illustration) are available only as captions and inline value labels in the extracted text; the visual Grad-CAM evidence in Section 5.3 could not be inspected directly.
- Numbers for the axial and sagittal crop variants are stated to be omitted from the current draft pending final log extraction, so the crop-orientation comparison is incomplete.
- A few Table 1 rows in the extracted text run together (for example the pet and vol group descriptions); the leakage-risk ordering was read from the surrounding prose.
- The paper is an arXiv preprint (arXiv:2609.15888v1); no code or trained-model release is stated, and ADNI data require an approved data-use agreement.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.