acceptodds
Withdrawn from ICLR 2027

Breaking the Atlas Barrier: Adaptive Supervoxels for Interpretable Alzheimer's Disease Diagnosis on PET

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia worldwide. Positron emission tomography (PET) reveals metabolic and molecular abnormalities whose location, extent, and spatial pattern vary substantially across individuals. Existing learning-based methods nevertheless aggregate evidence over predefined slices, patches, or atlas regions that are shared by all subjects. These fixed units misalign with individual abnormalities: they fragment coherent abnormal patterns or mix them with preserved tissue, which limits diagnostic accuracy and makes region-level explanations unreliable when abnormal uptake spans or falls between anatomical boundaries. We propose ASNet, an end-to-end framework that learns the spatial units themselves for PET-based AD diagnosis. Its Adaptive Supervoxel Partitioning (ASP) module learns differentiable voxel-to-supervoxel associations from PET features and spatial coordinates and iteratively refines them for each subject without any atlas. Its Abnormality Evidence Classifier (AEC) measures how far each learned region deviates from a cognitively normal reference and uses this deviation to route the region to an abnormality expert or a preservation expert. The class logits are the exact sum of the resulting region-level contributions, so the explanation is the diagnostic computation itself rather than a post hoc approximation. Across ADNI and AIBL on FDG, amyloid, and tau PET, ASNet attains the highest accuracy in all eight settings and an AUC of up to 96.2%. Substituting fixed atlases for the learned partition lowers AUC by over 10 points, and visual analyses confirm subject-specific, region-level evidence from the learned supervoxels.

Primary area:
applications to neuroscience & cognitive science
Keywords:
supervoxel, interpretable diagnosis, PET, atlas-free, neurodegenerative disease, Alzheimer's disease, medical image analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kJB5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper identifies fixed regions, patches, or slices as the bottleneck and proposes ASNet, which iteratively identifies subject-specific supervoxels that can localize the abnormalities better on PET scans using the ASP module. Then, it proposes an AEC module that can identify anomalous supervoxels using a reference to perform classification (AD vs CN and MCI vs CN) tasks on the ADNI and AIBL datasets. ASNet shows a good performance across the datasets and tasks it was evaluated for, although the dataset sizes were small.

Critical strengths

The paper proposes a supervoxel-based iterative partitioning technique on PET scans with task-specific adaptation. ASNet shows a good performance across 2 datasets and tasks, although the dataset sizes were small. However, I am currently unable to consider this a substantive strength for the reasons mentioned in the weaknesses and questions.

Critical weaknesses

1. Regarding the normative reference considered:

  • A normative reference in clinical neuroimaging is generally conditioned on age, which is not considered in this work. According to ASNet, the normative reference is obtained from all the CN subjects across all ages, which, to my knowledge, should not be called normative.
  • This also means that AEC characterizes the supervoxels using a common reference, which could be confounded by age, acquisition-related variations, etc. What if the CN cohort differs, or say new CN subjects from different cohorts get added for further training, or new test data from different scanners or age ranges are tested? I suspect the evaluations would change. There is no analysis of the framework's stability.
  • Also, with differences in brain volumes/structures across ages, and since the preprocessing performs only affine registration, there is no true correspondence across scans at the voxel level.

2. Potential overfiting: There is no mention of val split in the protocol, only 5-fold cross-val (meaning only train and test). How was the best ckpt chosen? This raises doubts of potential overfitting. Also, ADNI and AIBL provide longitudinal scans per subject, there is no mention of the train-test splitting strategy.

3. Missing details and experiments in the paper:

  • Missing training loss details. Sec 3.3 states that ASNet was trained in 3 stages, and mentions the use of distillation loss in stage 2, but the details of this loss is missing. What is getting distilled?
  • Missing architectural details, like the paper just mentions 3D convolution stem, MHSA, etc,. Details of the design are missing. There is no code to verify as well.
  • Missing generalizability test for cross-dataset.
  • Missing baselines that used supervoxels. SVGFormer, AdaSLIC, etc. are works that have used supervoxel for 3d medical volumes. The authors could consider either using their supervoxel partitioning approach or training these encoders for the classification tasks, or comparing the ASP with other established supervoxel-generation methods.
  • Missing ablations like using ASP’s K-means clustering fixed without the iterative update, random voxel selection or spatial groups without k-means, without abnormality gating, without reference, etc. These are necessary to understand the importance of the proposed framework.
  • No confounder analysis was performed, which is essential for understanding its effect on partitioning and, in turn, performance.

4. Regarding the claims:

  • Superiority over use of fixed atlas: AAGN uses HOA atlas in its framework and performs 95.19% AUC and 89.03% ACC. The proposed ASNet ablation of the same HOA atlas replacing ASP shows 85.86%AUC and 79.14%ACC. This shows quite a contrast of two things: a) its either the framework of ASNet does not support/handle atlas based inputs or the later part of the framework, AEC is overly fragile, b) The claim of superiority over atlas based grouping by 10% is misleading as the difference in AAGN vs ASNet is only ~1% in both AUC and ACC, thus, table 2 ablation cannot be considered for evaluating the atlas ablations.
  • Interpretable design by construction: the scores are not truly region-specific scores but rather coupled by the global features and hence have an influence from the global context. Also, the paper does not establish that the identified regions are truly anomalous, and hence, it's difficult to assess.
Questions for authors
  1. Regarding the motivation, if the abnormality lies on two parcels, we can consider both parcels to show anomalous behavior. If a well-trained model can identify both parcels as abnormal, why would that be considered unreliable? For example, say a scan has abnormality in hippocampus, at a later stage it spreads to the nearby parahippocampal region, then the model can flag that the hippocampus and parahippocampus shows abnormality, so what’s wrong here?

  2. How is the whole-brain context vector derived from the voxel features? What is the motivation to use FiLM? Because it is derived from the voxel features, it’ll just globally shift all the supervoxel representations in the same direction channel-wise. Another question here, if a scan is classified as say abnormal, then its not just certain supervoxels which are responsible, but also the global context, right? Then, how do I know what exactly was responsible? In other words, since each regional score depends on whole-brain context, its value need not originate exclusively from evidence inside the displayed region. How do the authors distinguish exact additive accounting from faithful spatial localization?

  3. In fig 4, I understand that the interpretations are different, but how do I read what is expected? Which methods failed? How do you claim that ASNet performs better? For example, even if say MAD-Former produces block-wise responses, are they pointing to the right region of abnormality? Also, the conclusion in Sec 4.4 is not entirely right. Similar to ASNet, AAGN also assigns a contribution to a parcel, which can be used to directly interpret its contribution for a classification. So, what is the additional interpretability property introduced by AEC.

  4. This also leads to my next question, do you have any ground truth to validate the regions identified to be abnormal? Or, in other words, how should I validate your qualitative images? So, I am not convinced by the evaluation of the partitions. The performance on tasks alone cannot support the partitioning technique.

  5. While the ASP uses a 3D conv stem, which I am assuming to be some k layers of CNN, ideally the output would not truly represent the exact voxel’s feature due to the design of kernel overlaps. Then is this assumption still valid? This matters because your assigned boundaries are wrt these voxel features, so we cannot assume that the boundaries are truly aligned to the corresponding voxels in the image space.

  6. When atlas was used, was it directly overlayed and ablated? Because mostly the atlases are designed for MNI152 1mm space. So, if it's different, then the overlaid regions will be misaligned, which could be the reason for poor performance rather than the strict atlas partitioning.

5 comments
AuthorsOpenReview ↗

Response to Reviewer kJB5 (Part 1/2)

[Part 1/2]

Main concerns

Weakness 1 (also the confounder analysis in Weakness 3). The CN reference in ASNet is our own operational definition and differs from a normative reference in clinical neuroimaging: it provides a weak prior on the PET image based on the CN subjects of the training split, which only decides whether a supervoxel is evaluated by the abnormal or the preserved expert, and does not by itself determine the diagnosis.

ASNet deliberately does not condition this reference on age, because there is no established age-specific uptake level that defines AD, MCI or CN. Like all compared methods, ASNet is trained on cohorts of mixed age without explicit age modelling. Regarding acquisition, PET values have different meanings for different tracers, so for an accurate interpretation each experiment is restricted to PET acquired with the same class of tracer.

ASNet does not require exact voxel correspondence across scans. Its comparison with the CN reference is made at the level of regions rather than individual voxels: each supervoxel is judged by its average deviation over all of its voxels, about 4,400 on average, so what matters is the overall uptake of a region, not exact voxel-to-voxel differences.

Weakness 4 (also Q6). The 85.86% AUC / 79.14% ACC is ASNet itself with ASP replaced by the fixed HOA partition; the 95.19% / 89.03% is AAGN, a different method designed end to end around atlas ROIs. The two numbers come from different models, so the contrast shows neither that ASNet cannot handle atlases nor that AEC is fragile: with only the partition fixed (Fixed K-Means), ASNet still reaches 93.93 AUC. The 10-point gap refers only to the atlas ablation within ASNet, not to atlas-based methods. We learn the partition instead of using fixed partitions such as atlases because such units are fixed in advance, so AD-related PET abnormalities, whose location and extent vary across individuals, may be fragmented by them or misaligned with them. ASP instead learns the partition dynamically for each subject under diagnostic supervision, so it can adapt to the different abnormality patterns of different subjects. To our knowledge, no existing PET-based AD diagnosis method learns such a subject-specific partition under diagnostic supervision.

The atlases are not misaligned: all PET images are registered to the SPM12 template (MNI) space and resampled to a 121×145×121 grid with 1.5 mm voxels, and AAL, BNA and HOA, defined in MNI space, were resampled to the same grid.

Q1. PET measures physiological uptake, and its abnormal patterns are not suitably explained with the concept of structural parcels. Abnormal areas are usually finer-grained than structural parcels and more patient-specific, so describing a structure as "abnormal" is not precise enough; their boundaries need to be fitted in more detail.

Q5. The assumption holds, and the boundaries stay tied to the image voxels. Every convolution in the stem has stride 1 and padding 1, with no pooling or downsampling, so the output feature map has exactly the same grid as the input image. Overlapping receptive fields make the feature of each voxel describe the voxel together with its immediate neighbourhood, as in any voxel-wise segmentation network, but do not move it to another voxel.

Q2 (also "interpretable by construction" in Weakness 4). The whole-brain context vector is a 16-d vector obtained as a learned weighted sum of the features of all brain voxels. FiLM is used because the uptake level of a PET region is partly related to the uptake of the whole brain and of its surroundings, AD-related changes are distributed, and clinical reading interprets a regional finding against the whole-brain pattern. A region should therefore be assessed with evidence from both the region itself and the whole brain, and requiring purely region-specific scores would not suit PET; this is by design.

Exact additive accounting and spatial localization are different properties, and the paper claims them separately. Exact additive accounting means that the image-level logit is exactly the sum of the supervoxel scores; this is what the paper means by contributions "faithful to the prediction", and it only concerns how the logit is decomposed over supervoxels. Spatial localization refers to the region boundaries, which ASP fits to the subject's PET at the voxel level, more finely than atlas parcels or patches. We do not claim that a region's score comes only from evidence inside its boundary: for the reason above, neither the pathology nor the model design is purely local.

AuthorsOpenReview ↗

Response to Reviewer kJB5 (Part 2/2)

[Part 2/2]

Experiments and other details

Weakness 2. There is a validation split: within each of the five folds, the data are split 3:1:1 into training, validation and test sets, and the best checkpoint is chosen by validation AUC; the test set is never used for training or checkpoint selection. For both ADNI and AIBL, only the baseline scan of each subject is used, and all splits are made at the subject level.

Weakness 3. What is distilled is the class posterior of the auxiliary classifier G-Cls: G-Cls is trained with ASP in stage 1 and then frozen as the teacher, and AEC is trained with a standard KL-divergence distillation loss to match its temperature-softened predictions, in addition to the cross-entropy loss.

The architecture is as follows:

  • Input and stem. Input 121×145×121. The stem has four 3×3×3 convolutions (stride 1, no downsampling) with InstanceNorm and GELU, and produces 16-channel voxel features.
  • ASP. 32-d supervoxel tokens; M=128M = 128 and T=2T = 2; candidate set NiN_{i} = the 10 spatially nearest centres; 4-head self-attention over the supervoxel tokens.
  • AEC. 16-d whole-brain context vector; abnormal and preserved experts are residual MLPs with LayerNorm, and the abnormal expert ends with softplus. A supervoxel is routed to the abnormal expert when its mean deviation exceeds τ=1\tau = 1 CN standard deviation.

Cross-dataset generalization is not a claim of the paper. Because PET intensities differ across tracers and acquisition protocols, ASNet is trained and evaluated within each cohort and tracer, where it achieves consistent gains on ADNI (FDG, AV45, AV1451) and AIBL.

The SVGFormer and AdaSLIC partitions and the K-Means clustering of ASP without the iterative update perform as follows in the same pipeline (M=128M = 128, FDG, AD vs. CN):

Partition (M=128M = 128)AUCACC
Fixed K-Means (ASP initialization, no iterative update)93.93±2.8286.45±3.94
SVGFormer partition94.55±2.5686.24±3.93
AdaSLIC partition90.75±5.7082.58±7.21
ASNet (learned)96.20±2.5390.54±2.09

The learned partition of ASP outperforms all of them. SVGFormer and AdaSLIC group voxels by image intensity, and Fixed K-Means by coordinates, all before training; ASP's assignment is learned with the classifier and updated with region context, so the regions are shaped by diagnostically relevant features.

Without the gate and the reference, AEC reduces to an unconstrained classifier on the same supervoxel representation, i.e. the G-Cls head, which Table 2 reports on the same ASP backbone: 94.14 AUC / 87.53 ACC, against 96.07 / 89.68 for AEC. Removing the monotonic prior tied to the reference, by letting the abnormal expert produce evidence for both classes, also lowers performance (FDG, AD vs. CN):

Abnormal expertAUCACC
Evidence for CN and AD (no monotonic prior)95.65±2.7888.82±3.44
Evidence for AD only (ASNet)96.20±2.5390.54±2.09

Q3 and Q4. ADNI and AIBL contain no voxel-level ground truth for PET abnormality, and none of the compared interpretable methods is evaluated against one. Explanations are therefore judged by their agreement with the image and with known AD patterns: a good explanation follows the uptake pattern of the subject's own image, covers the abnormal tissue without extending over preserved tissue, and emphasizes the most affected regions. In Fig. 4:

  • 3DSC-TF and MAD-Former respond in blocks that follow the patch grid; they point to the abnormality only roughly and do not provide a comprehensive explanation across the brain.
  • LA-GMF and LEAR highlight large areas, so the most affected regions do not stand out.
  • AAGN and PRGNN. Their explanation values change abruptly at parcel boundaries, which is inconsistent with the smooth disease patterns in brain images.
  • ASNet assigns its strongest AD evidence to the area of visibly reduced uptake in the upper part of the sagittal slice, with boundaries that follow the reduced-uptake pattern, consistent with the cortical hypometabolism typical of AD on FDG PET (Minoshima et al., 1995; Mosconi, 2005).

References.

  • Minoshima S. et al., A diagnostic approach in Alzheimer's disease using three-dimensional stereotactic surface projections of fluorine-18-FDG PET, J Nucl Med 36:1238–1248, 1995.

  • Mosconi L., Brain glucose metabolism in the early and specific diagnosis of Alzheimer's disease, Eur J Nucl Med Mol Imaging 32:486–510, 2005.

Thanks for the response. I have further questions:

Weakness 1:

  • The effect of confounding by age is not addressed. Even if the method does not consider age as input, the PET scans are confounded by age. So this is essential to understand its effect as you are considering CN reference irrespective of age.
  • Regarding correspondence, even if voxel correspondence is not required, since brains across ages or demographics can vary in sizes, there could be regional variations if the supervoxel identified is small. The authors also say that the boundaries need to fit the regions of abnormality in more detail. Is there any analysis conducted on the regions obtained and its correspondence to the reference?

Weakness 4:

  • If the 10% gap is with respect to ASNet itself and not wrt to atlas-based methods in general, then I feel you should clarify this in the paper, as it gives an impression that ASNet is better than atlas-based methods…

Q2: Could you provide ablations with and without this global feature?

Q5: Even if the grid size is the same, the receptive field varies with kernel size, overlapping kernels, and depth of the conv stem. So, feature location in the output does not truly correspond to the voxel in the input due to its local receptive field. The 4-layer stem creates a 9×9×99\times9\times9 receptive field, meaning each "voxel" feature already blends a wide neighborhood before supervoxel assignment. So, there is no reliable test to check how good the boundaries identified by the supervoxels are, as you claim that boundaries matter.

Weakness 2: Is it 3:1:1 split within each of the 5 folds? Please clarify the exact protocol in the paper.

Weakness 3:

  • It is suggested to present your final loss function used rather than as verbose as in the paper. Is there any ablation of this distillation loss?
  • The confounder analysis is still not addressed, which includes effects of manufacturer, demographics, etc. and its effects on partitioning.
Reviewer_ivZFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ASNet is a PET-based AD classifier that learns its own spatial units instead of using atlas regions, patches or slices. It starts from K-means centres on voxel coordinates (128 supervoxels) and refines the soft voxel-to-supervoxel assignments for two iterations, using feature similarity plus a spatial prior. Each supervoxel is then compared with a cognitively normal reference and sent to either an "abnormal" expert, which can only add evidence for AD, or a "preserved" expert. The regional contributions are summed into the logits, so the margin decomposes exactly over regions. Experiments use 5-fold CV on ADNI (FDG, AV45, AV1451) and AIBL (pooled amyloid) for AD vs CN and MCI vs CN, against nine baselines. The paper reports the best accuracy in all eight settings, AUC up to 96.2, and an AUC drop of more than 10 points when ASP is replaced by AAL, BNA or HOA. The interpretability evidence is qualitative, mostly single-subject figures.

Critical strengths

The idea is neat and the parts fit together: learn the regions, then make the classifier additive over them, so the explanation is the model's own computation. Gating against a normal reference with a non-negative abnormal expert is a sensible prior for PET. The evaluation covers three tracers and two cohorts, the reference is built from training folds only, and the paper is upfront about where it doesn't win (ADNI amyloid AUC). It is also clearly written.

Critical weaknesses

My main concern is that the central claim, that learning the partition is what helps, isn’t isolated. Table 2 swaps ASP for an atlas and AUC falls to 84–86. But AAGN and PRGNN in Table 1 use atlases and reach about 95 and 94 on the same task, so atlases don’t inherently cost ten points. The atlas variants probably lack something else ASP provides, like the voxel–region feature exchange and cross-region attention, and the paper doesn’t say how they were built. The control I’d want is the same pipeline on a fixed partition (the initial K-means supervoxels or SLIC). The T=1 result in Fig. 3 looks only a couple of points below T=2, which suggests a near-fixed grid may carry much of the performance. Most margins in Table 1 are 0.5–2 points against fold standard deviations of 2–5, with no significance tests, and T and M seem to be picked from the same CV results.

I’m also not convinced by the interpretability claims. An exact additive decomposition isn’t the same as localised evidence. Each supervoxel feature passes through attention across regions and a whole-brain context vector, so a region’s contribution isn’t a function of that region alone. There is no faithfulness test, and the evidence is a few single-subject figures. Since the partition is different for each subject, “supervoxel j” doesn’t mean the same place across subjects. That weakens the claim about cross-subject and longitudinal comparison, which rests on two AIBL subjects.

Questions for authors

Clarifications:

  1. How were the atlas variants in Table 2 implemented (same stem and feature exchange, only the assignment fixed)? Why do they score 84–86 AUC while AAGN and PRGNN score about 95 on the same task?

  2. How were T, M, τ, α and the stopping epoch chosen? If on the CV folds, can you rerun with nested CV or a held-out set? How are τ and the candidate set N_i defined?

  3. Are splits at subject level? Does Table 3 count subjects or scans, and how were CN subjects selected (85 for AD vs CN but 300 for MCI vs CN in tau)?

  4. Why are GF-Net and M3T missing for MCI vs CN? Is there a plain 3D CNN or PET-specific baseline (e.g. Rogeau et al.)?

  5. Sec. 4.2 says the partition "transfers across tracers and acquisition protocols", but each setting is trained separately. Is there a cross-cohort test?

Extra experiments:

  1. ASNet with a fixed K-means or SLIC partition at M=128, plus seeds and paired tests or bootstrap CIs for Tables 1–2.

  2. Deletion or insertion curves against Grad-CAM or IG on the same ASP backbone, and agreement with known AD regions.

3 comments
AuthorsOpenReview ↗

Response to Reviewer ivZF (Part 1/3)

[Part 1/3]

Main concerns

Weakness: learned partition. We learn the partition because AD-related PET abnormalities may span several brain regions or stay within a small localized area, and their location and extent vary substantially across individuals (Del Sole et al., 2008). Units fixed in advance and shared by all subjects (atlas regions, patches or slices) therefore split an abnormality or mix it with preserved tissue. ASP instead forms a new partition for every scan, including unseen scans at inference, from that scan's own features and coordinates, and the assignment is learned jointly with the classifier. Each patient's diagnosis and its explanation are thus expressed over regions shaped by that patient's own PET pattern. To our knowledge, no existing PET-based AD diagnosis method learns such a subject-specific partition under diagnostic supervision.

Weakness: atlas variants in Table 2 (also Q1). The 84–86 AUC in Table 2 is ASNet itself with its learned partition replaced by a fixed atlas partition (AAL, BNA or HOA). The roughly 95 and 94 AUC are AAGN and PRGNN, two different methods designed end to end around atlas ROIs. The two numbers therefore come from different models, and atlases do not inherently cost ten points: Table 2 shows only that the atlas ablation within ASNet loses more than 10 points. The atlas variants thus lose what ASNet is designed for: fixed partitions are fixed in advance and may fragment abnormalities or be misaligned with them, whereas ASP learns the partition dynamically for each subject under diagnostic supervision and can adapt to their different abnormality patterns.

The atlas variants change only the partition: the reassignment of voxels to regions is skipped, so every voxel stays in its atlas parcel throughout training, and the number of regions equals the number of parcels. Everything else in the network is unchanged, including the stem, the voxel–region feature exchange, the cross-region attention and AEC, and the parameter count is the same.

Weakness: cross-subject correspondence. This is intended: forcing the partitions of different subjects into one-to-one correspondence would bring back the population-level units that ASNet replaces. ASP adapts the partition to each subject, contributions are mapped back to voxels in the common template space, and explanations of different subjects are compared by location.

Weakness: localised evidence. Requiring the evidence to be purely local would not suit PET: PET measures physiological levels across the whole brain, AD-related changes appear as distributed patterns, and clinical reading interprets a regional finding against the whole-brain pattern. ASNet therefore expresses the evidence on subject-specific regions and uses the global modules as auxiliary context.

Weakness: faithfulness (also Q7). Each supervoxel's contribution is a term of the logit margin, not a post hoc attribution, so it is faithful to the prediction by construction.

References.

  • Del Sole A. et al., Individual cerebral metabolic deficits in Alzheimer's disease and amnestic mild cognitive impairment: an FDG PET study, Eur J Nucl Med Mol Imaging 35:1357–1366, 2008.
AuthorsOpenReview ↗

Response to Reviewer ivZF (Part 2/3)

[Part 2/3]

Experiments

Weakness: fixed-partition control (also Q6). The fixed partitions Fixed K-Means, SVGFormer and AdaSLIC perform as follows in the same pipeline (M=128M = 128, ADNI FDG, AD vs. CN):

Partition (M=128M = 128)AUCACC
Fixed K-Means (ASP initialization, no reassignment)93.93±2.8286.45±3.94
SVGFormer partition94.55±2.5686.24±3.93
AdaSLIC partition90.75±5.7082.58±7.21
ASNet (learned)96.20±2.5390.54±2.09

The dynamic partition of ASP outperforms all fixed partitions. SVGFormer and AdaSLIC group voxels by image intensity, and Fixed K-Means by coordinates, all before training; ASP's assignment is learned with the classifier and updated with region context, so the regions are shaped by diagnostically relevant features.

Weakness: T=1T = 1 vs. T=2T = 2. Performance for T=0T = 0–3 reassignments (M=128M = 128, ADNI FDG, AD vs. CN):

TAUCACC
0 (Fixed K-Means)93.93±2.8286.45±3.94
194.59±2.1487.10±2.04
2 (ASNet)96.20±2.5390.54±2.09
394.97±2.3787.10±4.46

T=0T = 0 is the fixed grid, i.e. Fixed K-Means. The performance requires a number of learned reassignments; a fixed grid is not sufficient.

Weakness: selection of TT and MM (also Q2). We explored T∈{1,2,3}T \in \lbrace 1, 2, 3 \rbrace (with M=128M = 128) and M∈{64,96,128,192,256}M \in \lbrace 64, 96, 128, 192, 256 \rbrace (with T=2T = 2), and the submission reports the best of these. We additionally re-selected TT and MM in each fold using validation AUC only: the resulting test AUC / ACC is 96.70 ± 1.83 / 90.75 ± 3.85, against 96.20 ± 2.53 / 90.54 ± 2.09 for the reported configuration, so the reported performance is not an optimistic result of test-based selection.

Protocol. Within each of the five folds, the data are split 3:1:1 into training, validation and test sets. The checkpoint is selected by validation AUC; the test set is never used for training or checkpoint selection. The CN reference uses only the training CN subjects.

Definitions.

  • τ\tau is fixed to 1 in all settings and not tuned; it represents one standard deviation of the CN reference.
  • α\alpha = 5 is the sharpness of the sigmoid that maps a supervoxel's deviation from the CN reference to its abnormality score; it is fixed in all settings and not tuned.
  • NiN_{i} is the set of the 10 supervoxels whose current centres are spatially nearest to voxel ii.

Weakness: significance (also Q6). The gains are stable across seeds and statistically significant.

Multiple seeds (ASNet and the four strongest baselines; the seed of Table 1 and two new seeds). For each method, the five folds are averaged within each seed, and we report mean ± std over the three seeds:

TaskAUC: ASNetAUC: best baselineACC: ASNetACC: best baseline
AD, FDG96.19±0.3595.17±0.34 (AAGN)90.53±0.1889.54±0.41 (LA-GMF)
AD, AV145194.10±0.0391.54±0.25 (AAGN)88.82±0.4885.10±1.11 (PRGNN)
AD, AV4592.51±0.1292.01±0.22 (AAGN)86.02±0.4785.56±0.22 (AAGN)
AD, AIBL93.05±0.0790.36±1.07 (LA-GMF)84.74±0.3283.65±0.45 (LA-GMF)
MCI, FDG67.47±0.1265.42±0.72 (3DSC-TF)64.68±0.1562.04±0.22 (3DSC-TF)
MCI, AV145173.14±0.1571.93±0.72 (3DSC-TF)68.00±0.3666.33±0.49 (LA-GMF)
MCI, AV4564.86±0.3265.15±0.35 (3DSC-TF)62.99±0.1261.97±0.67 (PRGNN)
MCI, AIBL73.06±0.0871.21±0.27 (3DSC-TF)68.92±0.2764.20±1.20 (3DSC-TF)

ASNet achieves the best results in most settings, and its seed-to-seed std is at most 0.48 points, well below its fold-to-fold std.

Significance testing. We compared ASNet with each of the nine baselines of Table 1 using a two-sided paired t-test over the 40 test folds (8 settings × 5 folds), with Holm correction over the nine baselines. The adjusted p-values are:

3DSC-TFAAGNGF-NetLA-GMFLEARM3TMAD-FormerMCENNPRGNN
AUC0.210.010<0.0010.001<0.001<0.001<0.001<0.0010.003
ACC0.0020.002<0.001<0.001<0.001<0.001<0.001<0.0010.002
AuthorsOpenReview ↗

Response to Reviewer ivZF (Part 3/3)

[Part 3/3]

Other questions

Q3. All splits are made at the subject level, and only the baseline scan of each subject is used, so Table 3 counts subjects, which equal scans. ASNet does not address class imbalance, so for AV45 and AV1451, where CN subjects greatly outnumber the patients, we selected as many CN subjects as there are patients in each task.

Q4. (a) GF-Net and M3T on MCI vs. CN. Both methods were proposed and evaluated only for AD vs. CN, which is why they were missing from the MCI half of Table 1. We have now trained both on MCI vs. CN in all four PET settings, with the same 5-fold splits and training setup as their AD vs. CN entries; ASNet outperforms both in AUC and ACC in every setting:

MCI vs. CN, AUC / ACC (%)ADNI FDGADNI AV1451ADNI AV45AIBL Amyloid
GF-Net53.06±9.35 / 59.46±3.6057.99±1.36 / 55.50±3.4857.67±4.60 / 53.08±3.3068.28±5.82 / 64.20±6.64
M3T50.94±5.37 / 51.53±7.5067.59±4.36 / 61.00±6.1656.64±3.30 / 51.54±2.4966.92±6.20 / 64.54±5.45
ASNet67.30±1.52 / 64.50±3.8173.35±4.64 / 68.50±3.7464.41±2.61 / 63.08±2.6272.94±4.90 / 68.59±4.44

(b) A plain 3D CNN. We added the PET-based 3D CNN of Rogeau et al. (2024) in all eight settings, using the same 5-fold splits as Table 1; ASNet outperforms it in AUC and ACC in all eight:

AUC / ACC (%)Rogeau et al. (3D CNN)ASNet
ADNI FDG, AD vs. CN93.73±2.70 / 87.53±4.3996.20±2.53 / 90.54±2.09
ADNI FDG, MCI vs. CN61.98±2.86 / 60.18±1.5567.30±1.52 / 64.50±3.81
ADNI AV1451, AD vs. CN78.75±2.09 / 50.00±0.0094.12±3.28 / 89.41±2.35
ADNI AV1451, MCI vs. CN69.16±3.60 / 62.83±3.4073.35±4.64 / 68.50±3.74
ADNI AV45, AD vs. CN89.65±3.43 / 83.06±3.8792.68±1.33 / 86.67±2.58
ADNI AV45, MCI vs. CN61.97±3.55 / 56.92±2.6464.41±2.61 / 63.08±2.62
AIBL Amyloid, AD vs. CN89.62±4.17 / 78.98±3.7593.15±2.82 / 85.19±3.51
AIBL Amyloid, MCI vs. CN66.98±6.53 / 62.20±7.3472.94±4.90 / 68.59±4.44

Q5. The sentence in Section 4.2 of the paper means that ASNet adapts to PET acquired with different tracers and protocols, not that a trained model transfers between them: trained within each setting, the same architecture achieves consistent gains on FDG, AV45 and AV1451 in ADNI and on AIBL, which pools three amyloid tracers.

References.

  • Rogeau A. et al., A 3D convolutional neural network to classify subjects as Alzheimer's disease, frontotemporal dementia or healthy controls using brain 18F-FDG PET, NeuroImage 288:120530, 2024.
Reviewer_TGt1OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ASNet for PET-based AD classification, which is an interpretable and subject-specific method. The ASNet has two modules. ASP learns subject-specific supervoxels, while AEC compares their uptake with a CN reference and assigns each region to an abnormality or preservation expert. The resulting regional contributions sum exactly to the class logits. The paper evaluates AD-vs-CN and MCI-versus-CN classification across three ADNI PET tracers and AIBL amyloid cohort. ASNet achieves the highest mean accuracy in all eight settings and the highest mean AUC in six. The paper also presents component ablations and hyperparameter sensitivity analyses with interpretable results.

Critical strengths
  1. The figure is nice and informative. I enjoy reading it. Also the final results and figures are easy to understand.
  2. The research question is useful and well motivated. Especially extending this to the subject level is meaningful. Based on the motivation, the method they built is suitable.
Critical weaknesses
  1. Because the AEC module uses the CN voxel density statistic as a reference, if I'm understanding correctly, they are assuming that AD voxel density statistic will be far away from the CN normal statistic distribution. However, i'm not sure if this is guaranteed in real world experiment. If the authors can show more evidence supporting this claim, it would strengthen the method.
  2. This is a minor weakness, as the methodology part is kind of hard to follow. Defining the method with a mathematical definition is great, but there should be some plain sentences motivating the components in the method.
Questions for authors

Could you provide some evidence regarding weakness 1?

1 comment
AuthorsOpenReview ↗

Response to Reviewer TGt1

Weakness 1 and the Question: evidence that AD deviates from the CN statistics.

  • Voxel-wise comparison with normal statistics separates AD from controls. Minoshima et al. (1995) built a normal database by averaging the scans of normal subjects, computed a Z-score for each patient at every surface pixel, and distinguished probable AD from normal subjects with 95–97% sensitivity at 100% specificity. Herholz et al. (2002) applied voxel-wise (age-adjusted) t statistics against 110 controls to 395 AD patients from eight centres, reaching 93% sensitivity and specificity for mild to moderate AD and 84% sensitivity at 93% specificity in very mild AD.
  • The pattern of deviation is well characterized. FDG PET shows reduced glucose metabolism in the parieto-temporal, frontal and posterior cingulate cortices as the hallmark of AD (Mosconi, 2005).

The reason is that PET measures a physiological quantity at each location: glucose metabolism for FDG, and amyloid or tau burden for the amyloid and tau tracers. After intensity normalization, this quantity is comparable across subjects, and AD appears precisely as its departure from the level of cognitively normal individuals.

Weakness 2: plain-language motivation for each component.

Adaptive Supervoxel Partitioning (ASP).

  • Why learn the regions. PET abnormalities do not follow atlas boundaries or patch grids, so fixed units split an abnormal area or mix it with preserved tissue. ASP therefore learns the regions for each subject.
  • Starting point. The brain is first divided into compact regions by K-means on voxel coordinates.
  • Reassignment. Each voxel is reassigned to a nearby region according to feature similarity and spatial distance, so voxels with similar uptake patterns end up together while regions stay compact.
  • Bidirectional update. Regions summarize their voxels and send this context back, so the next reassignment is informed by the regions; attention across regions lets each region take the others into account.

Abnormality Evidence Classifier (AEC).

  • Deviation from the CN reference. For each region, we measure how far its uptake lies from that of CN subjects, in units of the CN standard deviation.
  • Gate. Regions deviating by more than one CN standard deviation go to the abnormal expert, and the others to the preserved expert. Routing follows the measured deviation rather than a learned score, and is kept separate from the estimation of evidence.
  • Two experts. Abnormal and preserved tissue carry different diagnostic information, so each has its own expert. The abnormal expert can only add evidence for AD (non-negative output), which matches the clinical reading that a clear departure from normal uptake points towards disease.
  • Whole-brain context. The uptake level of a PET region is partly related to the uptake of the whole brain and of its surroundings, so each region is assessed with evidence from both the region itself and the whole brain, through a whole-brain context vector that modulates all regions.
  • Additive decision. The logit margin is the sum of the regional contributions, so the contribution map is exactly the computation behind the diagnosis.

References.

  • Minoshima S. et al., A diagnostic approach in Alzheimer's disease using three-dimensional stereotactic surface projections of fluorine-18-FDG PET, J Nucl Med 36:1238–1248, 1995.

  • Herholz K. et al., Discrimination between Alzheimer dementia and controls by automated analysis of multicenter FDG PET, NeuroImage 17:302–316, 2002.

  • Mosconi L., Brain glucose metabolism in the early and specific diagnosis of Alzheimer's disease, Eur J Nucl Med Mol Imaging 32:486–510, 2005.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.