List of Papers Browse by Subject Areas Author List
Abstract
Brain positron emission tomography (PET) has been widely used for the diagnosis of various neurodegenerative diseases. To assist physicians, convolutional neural networks (CNNs) and transformers have been explored for prediction of diseases based on brain PET images. While these models show promising performance, they are designed to process the entire image, which facilitates shortcut learning by extracting irrelevant features. To alleviate shortcut learning, we observe that brain images share the same structure, and regions of interest (ROIs) can be defined for relevant regions. In this regard, we propose Pyramidal Region Graph Neural Network (PRGNN), which employs a 3D convolutional backbone to learn multi-level feature representations and constructs nodes that correspond to anatomical ROIs. Using ROI-based node embeddings, PRGNN extracts metabolic patterns in functionally relevant regions and performs explicit inter-regional reasoning. We evaluate PRGNN on classifying 18F-fluorodeoxyglucose (FDG) and amyloid PET, outperforming models based on CNN, transformer, and GNN. Moreover, interpretability analyses highlight disease-relevant regions that align with clinical observations, demonstrating PRGNN’s potential for improving diagnostic performance and reliability. Code is available at https://github.com/Treeboy2762/PRGNN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0340_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Treeboy2762/PRGNN
Link to the Dataset(s)
BibTex
@InProceedings{KimDae_PRGNN_MICCAI2025,
author = { Kim, Daesung AND Seo, Seungbeom AND Kim, Boosung AND Choo, Kyobin AND Jun, Youngjun AND Yun, Mijin},
title = { { PRGNN: Pyramidal Region Graph Neural Network for Region-Based Brain PET Classification } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {554 -- 563}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel model, PRGNN, designed to enhance brain PET classification for neurodegenerative diseases by integrating 3D CNNs with graph neural networks (GNNs). The key idea is to use anatomical ROIs to generate graph nodes, capturing regional
- 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.
- A major strength of this paper lies in the innovative integration of a 3D CNN backbone with a graph neural network (GNN) architecture, enabling the model to effectively capture both local metabolic patterns and global anatomical relationships in brain PET images.
- The proposed PRGNN framework also excels in interpretability by quantifying the contribution of each anatomical region (node) to the final prediction, offering clinically meaningful insights.
- Furthermore, the paper includes a well-designed ablation study that systematically demonstrates the performance gains from incorporating graph convolutions at different stages of the CNN, validating the architectural choices and emphasizing the value of regional graph reasoning.
- 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 refers to neurodegenerative diseases such as AD (Alzheimer’s Disease), LBD (Lewy Body Dementia), and PSP (Progressive Supranuclear Palsy) in both the methodology and dataset descriptions. However, it does not provide definitions or clinical context for these terms, which limits accessibility and clarity for a multidisciplinary audience.
- The experiments rely primarily on a relatively small internal dataset and a subset of the ADNI database. It remains unclear why only 198 FDG PET scans from ADNI (137 NC and 61 AD) were selected, given the broader availability of relevant data in the ADNI repository. Additionally, important demographic information—such as mean age, sex distribution, and diagnostic subtypes—is missing, particularly for the internal dataset. These details are critical for evaluating potential data imbalance or bias.
- The authors mention merging the original 117 AAL atlas regions into 56 ROIs to reduce fragmentation (e.g., in the cerebellum), yet no criteria or validation are provided for this merging strategy. It is unclear whether this simplification may affect classification performance, and whether alternative merging schemes would lead to consistent results. A sensitivity analysis or ablation study addressing this decision would strengthen the methodological rigor.
- PET imaging is well-established for detecting early-stage Alzheimer’s pathology. Therefore, it is a missed opportunity that the authors did not include classification of earlier disease stages, such as Mild Cognitive Impairment (MCI). Including MCI in the classification task would significantly enhance the clinical utility of the proposed model, particularly for early intervention and prognosis.
- In Section 3.1, the interpretability analysis highlights key ROIs contributing to model predictions. While the authors assert clinical alignment, this section would be more convincing if it were supported with citations from prior literature confirming the disease relevance of the identified regions. Including such references would enhance the credibility and reproducibility of the interpretability findings.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While this paper introduces a novel and conceptually sound framework—PRGNN—for brain PET image classification using a hybrid 3D CNN and GNN architecture, and demonstrates its empirical advantages over several baselines, I recommend a weak reject due to multiple limitations that impact the work’s clarity, completeness, and broader scientific value.
The strengths of the paper include its innovative model design and strong emphasis on interpretability, with clear contributions in integrating regional graph reasoning and validating performance improvements through an ablation study. These aspects reflect thoughtful engineering and demonstrate promise for clinical applications.
However, the submission falls short in several key areas. First, the manuscript lacks definitions and clinical context for the diseases being classified (AD, LBD, PSP), which weakens accessibility for interdisciplinary readers. Second, the dataset description is incomplete and insufficiently transparent—specifically regarding the rationale behind the subset of ADNI scans used, and the absence of demographic information needed to assess potential bias or imbalance. Third, a critical preprocessing step involving ROI merging is not justified or validated, raising concerns about reproducibility and robustness. The paper also misses the opportunity to include earlier disease stages, such as MCI, which are crucial for early intervention and highly relevant to PET-based diagnosis. Finally, although the interpretability analysis is a highlight, it lacks supporting citations from clinical literature to substantiate the disease relevance of the identified ROIs.
- 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.
The authors have provided a reasonable response to my previous comments.
Review #2
- Please describe the contribution of the paper
This paper proposed a unified network for combining multi-level features from CNN and spatial correlations from GNN for disease diagnosis.
- 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 multi-stage CNN provides a multie-level feature extraction strategy.
- The GNN network provides a solution for considering the spatial coherency among ROIs.
- The results are interpertable, providing the importance ranking among ROIs in diagnosis.
- 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.
-
Lack of novelty in model architecture: The use of CNNs for feature extraction and the integration of CNN and GNN components is not novel. For example, citation [9] already demonstrates this architectural combination, and the current paper does not introduce meaningful architectural innovations beyond existing frameworks.
-
Invalid downstream task setup: Using FBB PET images to classify amyloid positivity is problematic, as the label (amyloid status) is directly derived from the same imaging modality. This introduces a circular logic that undermines the validity of the task. Furthermore, even for this relatively straightforward classification, the proposed method does not show significant performance improvement over AAGN, despite having a much larger model size.
-
Missing abbreviation definitions: Several abbreviations (e.g., LBD, PSP) are used without proper definitions. These should be clearly introduced at first mention to ensure accessibility for readers not familiar with these clinical terms.
-
Ambiguity in notation: The symbol X is reused in different contexts across the paper, referring both to the input 3D PET volume and the node features in the graph (Section 2.3). This overlapping notation introduces confusion and should be revised for clarity and consistency.
-
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method lacks sufficient novelty compared to prior approaches that have been applied to other imaging modalities. Additionally, the experimental evaluation is limited and does not adequately validate the effectiveness or generalizability of the model.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper introduces the Pyramidal Region Graph Neural Network (PRGNN), which integrates a 3D CNN backbone with a region-based graph neural network. PRGNN defines graph nodes using anatomically relevant regions of interest (ROIs) in brain PET images(rather than the entire voxel grid), thereby constraining the model to focus only on functionally significant brain areas. The proposed approach leverages the advantages of both 3D CNNs and GNNs to combine fine-grained local patterns with high-level structural context, improving classification performance. Additionally, it aggregates the logit contributions from each node (or ROI), providing an interpretable prediction.
- 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.
- This paper provides a clear and detailed description of the proposed method by breaking it down into separate steps.
- The overall model architecture is very well illustrated in Figure 1, which greatly helps readers understand the approach.
- By employing a 3D CNN backbone for brain PET images and using clinically important anatomical regions as nodes, we further advanced the previous method [1].
- In Section 2.2, using only valid ROIs as nodes, the approach alleviates the unnecessary background and shortcut learning issues that can occur in typical CNN and Transformer models.
- By reporting the contribution of each ROI in Table 3, it explicitly shows how much each brain region influences the final prediction, enabling clinical interpretation of the results. [1] Kai Han, Yunhe Wang, Jianyuan Guo, Yehui Tang, and Enhua Wu. Vision gnn: An image is worth graph of nodes. Advances in neural information processing systems, 35:8291–8303, 2022.
- 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 mentions that some of the 117 regions in the AAL atlas were merged to simplify the analysis, but it should specify the criteria and methods used for this merging process.
- When constructing the graph by connecting nodes, the paper uses K-Nearest Neighbor, but it is unclear on what basis K was chosen, and whether experiments were conducted to analyze the results with different values of K.
- The evaluation section should include a more detailed description of how the k-fold cross-validation was conducted.
- In the Abstract, "inadvertantly" should be corrected to "inadvertently" and "anatomic ROIs" should be changed to "anatomical ROIs." - In all Table and Figure captions, a period (.) is missing at the end of the final sentence and should be added. - For Table 1, the content in each cell should be spaced out adequately for improved readability. - 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper introduces PRGNN integrating a 3D CNN and GNN for PET image analysis, enabling more refined analysis. By treating brain regions as nodes and using node embeddings, the method effectively addresses issues of unnecessary background inclusion and shortcut learning present in previous approaches. Additionally, the extraction of contribution scores from each brain region, along with Grad-CAM visualizations, provides an interpretable link between the regions and the final prediction. However, the paper lacks detailed descriptions of the criteria and methods used for merging regions, the exact procedure for the k-fold cross-validation, and the rationale behind the selection of the hyperparameter K. Addressing these issues would make the approach even more compelling and could potentially boost its overall evaluation to an Accept recommendation.
- 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.
The authors properly addresssed weakness points raised by the reviewers. The paper seems acceptable if the suggested clarifications are incorporated into the revised manuscript.
Author Feedback
We sincerely appreciate the detailed feedback. We will incorporate the suggested clarifications into the final version.
R1/R3 Abbreviations and definitions: We clarify that Alzheimer’s Disease (AD), Lewy Body Dementia (LBD), and Progressive Supranuclear Palsy (PSP) are distinct neurodegenerative diseases characterized respectively by amyloid/tau pathology, α-synuclein inclusions, and 4R-tauopathy; diagnostic criteria for each disease have been cited in [22-24]. For clinical context, the diseases being classified (AD, LBD, PSP) are representative neurodegenerative diseases that exhibit distinct patterns in FDG PET.
R1 Dataset details: A critical criterion that limited our selection was amyloid scan outcome for AD diagnosis, which left only 72 subjects with confirmed amyloid positivity. To address concerns about bias, we report the demographics of our internal dataset: the mean age ± SD (in years) was NC 62 ± 15; AD 71 ± 8; LBD 66 ± 9; PSP 68 ± 6. The sex distribution (in % female) was NC 62%; AD 69%; LBD 35%; PSP 38%. Among 71 AD cases, 42 were MCI and 29 were dementia.
R1/R2 Merging of AAL atlas: The merging scheme was designed to improve classification performance by minimizing redundancy while maintaining clinical coherence, similar to (Nozadi, Int. J. Biomed. Imaging, 2018). Specifically, we observed overfitting when all original 117 AAL atlas regions were used for training, likely due to small ROIs introducing noisy features that hinder generalization. For validation, different merging schemes can be designed, such as not merging the medial temporal lobe. We merged AAL labels along gyral boundaries, removing arbitrary anterior-posterior splits within the same gyrus. For example, we unified the medial, dorsolateral, and orbital subdivisions of the superior frontal gyrus because they share a single principal sulcus and underlying cytoarchitecture. Next, we grouped fragmented areas, such as the 9 fragments of the left cerebellum. For reproducibility, we will release the merged AAL atlas and complete mapping dictionaries.
R1 Missing MCI classification: Our focus of the study was on the differential diagnosis among neurodegenerative diseases, not staging of the diseases. Specifically, as AD-MCI and AD-dementia differ only in the severity of cognitive and functional impairments, we labeled them as AD altogether.
R1 References for key ROIs: We acknowledge the need for supporting citations, which we cite as follows: AD, thalamus (De Jong, Brain, 2008) and sup. parietal (Jacobs, Neurosci Biobehav Rev., 2012); LBD, lingual and cuneus (Whitwell, J. Nucl. Med., 2017); PSP, putamen (Meyer, J. Nucl. Med., 2017)
R2 Hyperparameter tuning of K: We tuned K from 3 to 12 empirically and observed the best performance when K=9.
R2 K-Fold Cross-validation: We followed standard 5-fold cross-validation setup, but performed an additional stratified 8:2 train/val split within each training set.
R3 Lack of novelty: AAGN [9] is not a GNN architecture, for it uses a gating network to assign weights to ROI features. This mechanism does not model inter-regional relationships, which is necessary for classifying complex scans like FDG PET. In contrast, PRGNN introduces a novel pyramidal node embedding mechanism for graph convolution, enabling inter-regional reasoning at different levels of abstraction. Furthermore, our model performs pooling across channels for direct region-level interpretability, which was infeasible in previous CNN-GNN-based models.
R3 Invalid downstream task setup: While the amyloid status is derived from FBB PET images, assessment of amyloid positivity requires either experienced readers or a structural MRI. Our downstream task is designed to replicate this assessment, as demonstrated by previous studies (e.g., Fan et al., Radiology, 2024). Our model exhibited the best accuracy on average and has added value in evaluating regional contributions.
R3 Notation: We will denote node features with the symbol E for consistency.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper combines CNN and GNN for AD classification based on PET images. R1 (Very confident) and R3 suggest weak rejections, while R2 suggests acceptance after rebuttal. After reading the review comments and rebuttal, several concerns remain on the technical novelty, limited accuracy improvement, and problematic task design (MCI classification would be clinically more useful, but was ignored; the amyloid status is derived from FBB PET image).
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A