List of Papers Browse by Subject Areas Author List
Abstract
Accurate and computationally efficient diagnosis of Alzheimer’s Disease (AD) from MRI remains a challenge in medical imaging.
While 3D deep learning models capture volumetric biomarkers, they require substantial GPU memory. Conversely, 2D approaches are efficient but fail to capture spatial context, limiting their clinical applicability. We present ADAPT, a clinically-inspired diagnostic framework that achieves 3D-level accuracy with 2D computational costs through careful co-design of three innovations:
(1) a View-specific transformer architecture with four encoder families that separately learn within-slice, within-view, and cross-view relationships;
(2) an adaptive view-profiling strategy that dynamically allocates more slices to medically informative views based on learned attention scores, providing more context in limited burden; and
(3) a pathology-driven morphology augmentation that simulates progressive atrophy patterns to improve model robustness.
Experiments across various AD datasets demonstrate that ADAPT achieves improved performance over CNN and transformer baselines while costing \textbf{75.8\%} lower GFLOPs. Beyond standard AD vs.\ NC classification, ADAPT also achieves strong performance on the \emph{early-stage} task of NC vs.\ MCI, further supporting its relevance for real-world application. These results highlight ADAPT as a practical, interpretable, and computationally efficient alternative to previous models for cross-institutional Alzheimer’s Disease diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5721_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYif_ADAPT_MICCAI2026,
author = { Wang, Yifeng AND Chen, Ke AND Wang, Haohan},
title = { { ADAPT: Adaptive Profiling Transformers for Efficient Alzheimer’s Disease Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes ADAPT, a framework for Alzheimer’s disease diagnosis from MRI that aims to achieve 3D-level diagnostic performance while maintaining the efficiency of 2D models. The method introduces a view-specific transformer architecture that processes MRI slices from different anatomical views and models relationships at multiple levels, including within-slice, within-view, and cross-view dependencies. In addition, the framework incorporates an adaptive view-profiling mechanism that dynamically allocates more slices to diagnostically informative views based on learned attention scores. The paper also introduces a pathology-driven morphology augmentation strategy designed to simulate progressive brain atrophy patterns associated with Alzheimer’s disease. Through these components, the proposed framework seeks to balance diagnostic accuracy, computational efficiency, and clinical relevance, demonstrating improved performance on AD classification tasks (including AD vs. NC and NC vs. MCI) across multiple datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Authors introduce an adaptive view-profiling mechanism that dynamically allocates slices to diagnostically informative views based on learned attention scores. This is a thoughtful design choice because it attempts to prioritize clinically relevant anatomical regions while maintaining computational efficiency, potentially allowing the model to extract richer contextual information without processing the full 3D volume. The proposed pathology-driven morphology augmentation strategy is another interesting attempt to incorporate disease progression knowledge into the training process by simulating atrophy-related morphological variations along the NC–MCI–AD continuum. Finally, the paper presents empirical evaluation across multiple AD classification tasks while also reporting substantial reductions in computational cost compared to other models.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
First, the experimental design relies primarily on training and testing within the ADNI dataset, which may not adequately evaluate the model’s ability to generalize across acquisition protocols and study cohorts. Since ADNI contains multiple phases with different scanners and imaging protocols, it would be more informative to treat ADNI-1, ADNI-2, and ADNI-3 as separate domains for training and testing, which would better reflect real-world cross-cohort generalization. Second, the manuscript does not report the sample size and class distribution for the additional datasets used in the experiments (AIBL, MIRIAD, and OASIS). Without this information, it is difficult to interpret the reported performance or assess whether the datasets are balanced, which is important for judging the reliability of the accuracy results. Third, the evaluation focuses mainly on binary classification tasks (e. g. , AD vs. NC or NC vs. MCI), whereas a three-class classification setting (AD vs. MCI vs. NC) would be more clinically meaningful and more consistent with the diagnostic continuum of Alzheimer’s disease.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on the paper’s attempt to address an important problem in Alzheimer’s disease diagnosis from MRI, namely the trade-off between capturing 3D anatomical context and maintaining computational efficiency. The proposed ADAPT framework introduces a multi-view transformer architecture with view-specific encoders, an adaptive view-profiling strategy, and a pathology-driven morphology augmentation mechanism to approximate volumetric understanding while keeping the computational cost close to that of 2D models. This design is practically appealing because it aims to provide 3D-level diagnostic performance with significantly lower computational requirements, which could improve the feasibility of deploying such models in real clinical environments with limited resources. The empirical results also show competitive performance across multiple AD classification tasks and datasets while substantially reducing GFLOPs compared with conventional 3D approaches, suggesting potential practical value.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Authors feedback satisfactory.
Review #2
- Please describe the contribution of the paper
The main contribution of the paper is the proposed ADAPT framework for Alzheimer’s Disease diagnosis from MRI, which achieves 3D-level diagnostic performance using a computationally efficient 2D-based design. It introduces a view-specific transformer architecture to model within-slice, within-view, and cross-view relationships, along with an adaptive view-profiling strategy that allocates more informative slices based on learned attention. Additionally, it incorporates a pathology-driven morphology augmentation technique to simulate disease progression and improve robustness.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has several notable strengths.
First, it presents a well-motivated framework that bridges the gap between 2D efficiency and 3D volumetric modeling in Alzheimer’s disease diagnosis. The idea of achieving 3D-level performance using a computationally efficient 2D-based design is practically important, especially for real clinical settings with limited GPU resources.
Second, the view-specific transformer architecture is a meaningful contribution, as it explicitly separates within-slice, within-view, and cross-view relationships. This structured modeling of multi-view MRI data is a reasonable way to capture spatial dependencies that are often missed in standard 2D approaches.
Third, the adaptive view-profiling strategy adds an intelligent mechanism to prioritize more informative slices based on learned attention, which aligns well with the clinical notion that not all slices contribute equally to diagnosis.
Finally, the pathology-driven morphology augmentation is a useful data-centric contribution that helps improve robustness by simulating disease progression patterns, and the reported results across multiple AD-related classification tasks suggest consistent performance improvements over baseline methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper has several limitations.
First, some components are related to prior work on multi-view and slice-based transformer/CNN models for MRI analysis, such as multi-view CNNs and slice-attention methods used in Alzheimer’s disease classification, but the comparison with these methods is not clearly discussed.
Second, the efficiency claims are mainly supported by GFLOPs reduction, without detailed evaluation of real-world latency or memory usage in clinical settings.
Third, similar pathology-aware augmentation strategies have been explored in earlier work on brain atrophy simulation and disease progression modeling, but the paper does not clearly compare against them.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend an accept. The paper presents a well-designed and clinically motivated framework for Alzheimer’s disease diagnosis that effectively balances accuracy and computational efficiency. The proposed ADAPT model introduces a structured way to capture within-slice, within-view, and cross-view relationships using a view-specific transformer, which is appropriate for modeling MRI data. The adaptive view-profiling strategy and pathology-driven augmentation further strengthen the method by improving the model’s ability to focus on informative regions and handle disease variability. The experimental results show consistent improvements over CNN and transformer baselines across multiple datasets and tasks, including both AD vs NC and NC vs MCI classification. Additionally, the significant reduction in computational cost while maintaining strong performance makes the approach practical for real-world clinical deployment. Overall, the combination of methodological design, empirical performance, and efficiency considerations supports the acceptance decision.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Accept
Review #3
- Please describe the contribution of the paper
This paper presents ADAPT, a clinically-inspired diagnostic framework that achieves 3D-level accuracy with 2D computational costs through careful co-design of three innovations.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
AD diagnosis is crucial task.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.It is better to explain the acronyms at first occurance. 2.Reviewer is not sure about the relationship between 3 parts in Figure 1.It is recommanded to use same color for same concept (e.g., all use blue for “embedding”). 3.Reviewer understands page width is narrow, but Table 1 and Table 2 are very hard to read. 4.In Table 1-3, best in each column should be highlighted. 5.Layout of Figure 2 and 3 are not good, Figure 3 is never mentioned, they could be combined into 1 figure (and 2 subfigures). 6.In Figure 2 and 3, the role of each column is not clearly written in both figures and Section 3.3.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The overall presentation of this paper is satisfactory, but flowchart and some experiments are unclear.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and AC for recognizing ADAPT’s clinical motivation, view-profiling design, pathology-driven augmentation, and consistent gains. We address the concerns below: 1) Cross-cohort setup and cohort reporting (R1, AC). Our robustness claim is based on external-cohort testing, not an ADNI-only evaluation. ADNI was used for training/validation/model selection; AIBL, MIRIAD, and OASIS were held-out test cohorts never used for model selection. ADNI-1/2/3 phase-wise splitting is useful but remains intra-consortium; testing on independent studies/institutions more directly supports cross-institutional generalization. We will report external cohort composition: AIBL 413 scans (363 NC/50 AD), MIRIAD 523 (177 NC/346 AD), and OASIS 2157 (1692 NC/465 AD). Reliability comes from subject-level disjoint ADNI splits, model selection only on ADNI validation data, and consistent gains across external cohorts with different sizes/class balances, reducing the chance that improvement is due to one favorable test set. For three-class AD/MCI/NC, our binary tasks cover AD vs. NC for standard separation and NC vs. MCI for early-stage discrimination, where differences are subtler. Joint three-class evaluation is valuable but complementary to our main claim. 2) Efficiency beyond GFLOPs (R2, AC). We profiled inference under matched hardware settings. ADAPT runs at ~52 ms/volume with only 0.43 GiB peak GPU memory. Compared with the closest multi-view baseline M3T, ADAPT is ~4.9× faster (52 vs. 254 ms), uses ~56× less memory (0.43 vs. 24.3 GiB), and requires ~1/10 GFLOPs (92.5 vs. 934.4), while achieving higher accuracy across all cohorts in Table 1.Compared with nnMamba, ADAPT is slower in latency (52 vs. 25 ms) but uses ~6× less memory and ~5× fewer GFLOPs with higher accuracy. Thus, our claim is a favorable accuracy-memory-complexity trade-off, not merely GFLOPs reduction. We will add the numbers to Table 1.3) Positioning against prior work (R2, AC). ADAPT is related to multi-view/slice-attention MRI models, but differs mechanistically. Compared with M3T [8] and MTr-Net [21], prior multi-view/2.5D methods use fixed plane processing and do not learn to reallocate slice budget across multi-views according to disease relevance. ADAPT couples view-specialized encoders with learned adaptive slice allocation while keeping total computation fixed. Under the same morphology-augmentation setting, ADAPT outperforms M3T across all four cohorts with ~1/10 GFLOPs. Compared with MSA3D (Chen et al., 2022), which uses attention to weight/fuse slice features while retaining a 3D global-feature branch, ADAPT changes which slices are sampled and uses a 2.5D transformer without full-volume 3D convolution. For pathology-aware augmentation, we claim that the morphology perturbation is only a training-time regularizer to increase NC-MCI-AD morphological variability. 4) Presentation and reproducibility revisions (R3, AC). R3’s concerns mainly involve readability/organization rather than method/result validity. We will revise Fig. 1 to make the pipeline explicit: MRI slicing → SAE shared encoding → DS-AE view-specific encoding → IntraCAE within-view fusion → InterCAE cross-view fusion → classification/adaptive profiling, with consistent colors for embeddings, attention blocks, and view streams. We will bold the best values in Tables 1-3, reformat Tables 1-2, and combine Figs. 2-3 into one NC-vs-AD attention visualization with row/column labels (views/encoder-stage attention maps) and explicit reference in Sec. 3.3.We will also define acronyms at first occurrence, release code upon acceptance, and report missing details including hardware, preprocessing, and latency/memory protocol. These revisions directly address R3 and do not alter the method, experiments, or conclusions.
Chen, L., Qiao, H., & Zhu, F. (2022). Alzheimer’s disease diagnosis with brain structural MRI using multiview-slice attention and 3D convolution neural network.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
- The paper proposes a clinically motivated ADAPT framework that aims to achieve 3D-level AD diagnosis performance with 2D-level computational cost. The view-specific transformer, adaptive view profiling, and pathology-driven morphology augmentation are well aligned with the goal of balancing accuracy, efficiency, and clinical relevance.
- The evaluation relies heavily on ADNI, and the paper does not clearly report sample sizes and class distributions for the additional datasets. A clearer cross-cohort setup, such as treating ADNI phases or external cohorts as separate domains, would better support the claimed robustness.
- The positioning against prior work needs clarification. Several components overlap with existing multi-view, slice-attention, and pathology-aware augmentation methods, but the paper does not sufficiently compare with or discuss these related approaches. This makes it harder to judge the true methodological novelty.
- The efficiency and experimental claims should be better supported. GFLOPs reduction alone is not enough to demonstrate practical efficiency; latency, memory usage, and clearer reporting of tables/figures and experimental settings are needed.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal satisfactorily addresses the main concerns by clarifying the held-out external cohort setup, reporting cohort composition, adding latency and memory evidence beyond GFLOPs, distinguishing ADAPT from prior multi-view/slice-attention and augmentation methods, and resolving presentation/reproducibility issues. Therefore, I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Following the rebuttal, all three reviewers agreed to accept the paper, including R3, who had initially rejected it. The authors credibly addressed the main concerns by clarifying the cross-cohort setup and showing the efficiency claim using latency and memory measurements rather than GFLOPs alone. Overall, the methodological design, external-cohort validation and favourable accuracy-efficiency trade-off justify acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper addresses an important and clinically relevant problem: how to retain strong 3D diagnostic performance for Alzheimer’s disease classification while substantially reducing computational cost. I find the overall framework well motivated, and the three main components of ADAPT, namely the view-specific transformer design, adaptive view profiling, and pathology-driven morphology augmentation, are well aligned with this goal.
The rebuttal adequately addresses the main concerns raised in the reviews and meta-review. In particular, the authors clarify that the evaluation is not limited to ADNI-only testing: ADNI is used for training/validation/model selection, while AIBL, MIRIAD, and OASIS are treated as held-out external cohorts. This substantially strengthens the robustness claim. The rebuttal also improves the efficiency argument by reporting latency and peak GPU memory in addition to GFLOPs, which makes the practical trade-off much clearer and more convincing.
I also appreciate the improved positioning relative to prior multi-view and slice-based methods. While the individual components are related to existing ideas, the paper’s contribution lies in a coherent and practically meaningful integration: learning how to allocate limited slice budget across views while maintaining a lightweight 2D/2.5D design. In this sense, the work is not purely incremental engineering, but a thoughtful methodological refinement with demonstrated empirical value.
There are still some limitations. The novelty is moderate rather than radical, and some important implementation and reporting details should have been clearer in the original manuscript rather than only in the rebuttal. A three-class AD/MCI/NC setting would also have strengthened the clinical scope. However, I do not view these as fatal weaknesses. The paper presents a solid technical contribution, a strong efficiency-accuracy trade-off, and encouraging external-cohort results that are relevant to the MICCAI community.
Overall, I believe the paper is sufficiently strong for acceptance.