List of Papers Browse by Subject Areas Author List
Abstract
Multi-site functional MRI (fMRI) is essential for learning generalizable functional connectivity (FC) biomarkers, yet scanner- and protocol-specific variability shifts FC distributions and degrades cross-site generalization, necessitating effective harmonization. Existing methods often rely on simple many-to-one global mappings toward a single pre-defined target site and are also difficult to transfer across atlases. To address these issues, we propose UniFCHarm, a unified prompt-guided framework for multi-scanner and multi-atlas FC harmonization. UniFCHarm first pretrains a GraphCLIP using 44,493 acquisition- and atlas-aware FC-prompt pairs to align them into a shared latent space, enabling harmonization toward a user-specified target condition instead of being constrained to a single reference site or atlas. It then performs coarse-to-fine harmonization by first correcting subnetwork-to-subnetwork connectivity and subsequently refining ROI-to-ROI connections, thereby mitigating over-correction. Evaluations on traveling-subject benchmark across multiple atlases demonstrate that UniFCHarm achieves the highest ground-truth agreement with MAE 0.033 ± 0.001 and reliably matches the target acquisition setting with accuracy 91.25 ± 6.18%, outperforming the evaluated representative baselines. Our code is available at: https://github.com/yizhixiaolinya/Code-MICCAI2026_fMRI-Harm.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1169_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{LinXin_UniFCHarm_MICCAI2026,
author = { Lin, Xin AND Liu, Yuxiao AND Wang, Yulin AND Teng, Lin AND Yang, Long AND Zhao, Shilun AND Zhang, Kai AND Cui, Zhiming AND Shen, Dinggang},
title = { { UniFCHarm: Unified Prompt-Guided Harmonization of fMRI Functional Connectivity Across Scanners and Atlases } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this submission, the authors introduce a new method for harmonizing functional connectomes. Their approach achieves prompt-guided harmonization by aligning the outputs of a graph encoder and a text encoder. The harmonized graphs are reconstructed in two steps: the method first generates a network-level prediction, which is then refined for the ROIs of the 7 networks in the Yeo atlas.
- 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.
. As far as I know, prompt-guided connectome harmonization is a novel idea. Whether novel or not, it is definitely a promising approach. . The method can handle multiple atlases / connectome resolutions, which is very interesting for applications . Strong method validation, involving a large data set (44,493 connectomes from 14 databases) and a comparison with 4 SOTA alternatives . The manuscript is well written and easy to follow.
- 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.
major weakness
. The need for a coarse-to-fine strategy to generate the harmonized connectomes is a main weakness of the approach. This issue required a more complex approach, and raises questions about the atlas used for the refinement: would the authors get better results with a different atlas, particularly if there were greater similarity between the atlas and the target connectome atlases? Since additional experiments cannot be conducted to assess this point during the rebuttal phase, I would suggest mentioning this issue as a method limitation in the conclusion.
minor weaknesses
. The author should provide more details about the datasets and the implemented methods. For instance, the most important dataset for the validation, the SRPBS traveling-subject benchmark used to derive the entire Table 1, is not described at all. How many scans does it contain? The downstream analysis is not described either: which classifier was implemented? How many NC/MCI/AD scans were used?
. The authors should report the training time, not just the number of epochs and hardware settings.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
Despite minor weaknesses, this work is of excellent value and would be worth presenting at MICCAI. I would suggest publication/presentation once the authors have provided the few missing details I mentioned.
- 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 addressed my comments, so I am confident the paper will be in good shape for publication.
Review #2
- Please describe the contribution of the paper
This manuscript proposes a prompt-guided framework for multi-scanner and multi-atlas functional connectivity harmonization, aiming to improve brain fMRI-based classification. The method is evaluated on the SRPBS dataset.
- 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 proposed method consistently outperforms the selected baseline methods across the reported evaluation metrics. An ablation study is also conducted to assess the contribution of individual components within the framework.
- 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 motivation for the proposed framework is not sufficiently justified. It remains unclear why this specific model design is appropriate for the problem.
There are important missing technical details. For example, no loss function is explicitly defined in Section 2.1 (GraphCLIP pre-training for prompt–FC alignment). Key variables, such as N_a, N_c, d, and k, are introduced without clear definitions. The descriptions of the text and graph encoders are also insufficient.
The manuscript lacks essential details required for experiments. In particular, it does not clearly describe how the models, including baseline methods, were trained, how hyperparameters were selected (e.g., the choice of lambda=0.2), or which fMRI preprocessing pipeline was used. The data information is insufficiently described.
It is also unclear how the ground-truth functional connectivity (FC) is defined in this study.
The selection of baseline methods (e.g., [3][15][26]) is not sufficiently justified, and the relationships between these methods and the proposed approach are not clearly explained. Furthermore, several related works discussed in the Introduction (e.g., [8][19][21][25]) are not included in the experimental comparisons without justification.
- Please rate the clarity and organization of this paper
Poor
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Technical details of the proposed framework are insufficiently described, and several notations are not clearly defined. The motivation for the method is also not fully justified. In addition, important experimental details, such as training procedures, hyperparameter selection, and preprocessing steps, are missing, raising concerns about reproducibility and fairness of comparisons. While the method shows empirical improvements and includes ablation studies, these issues limit confidence in the results.
- 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.
I have read the other reviews and the authors’ rebuttal. The authors addressed my major concerns. I recommend acceptance, provided the clarifications and additional details promised in the rebuttal are incorporated into the camera‑ready version.
Review #3
- Please describe the contribution of the paper
1) This work presents UniFCHarm, a unified prompt-driven framework for the harmonization of functional connectivity (FC) that works under different scanning sites and brain atlases. 2) The approach uses pretrained GraphCLIP to map structured prompts that include the information of acquisition sites and brain atlas into the latent space of FC graphs, facilitating a flexible prompt-driven process for harmonization towards any desired target domain. 3) A hierarchical coarsening-to-refining harmonization strategy to harmonize connectivity at both subnetwork-level and ROI-level, tackling multi-scale effects during the harmonization process while minimizing over-harmonization was also incorporated.
- 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.
1) Successfully deals with variability of scanners and atlases, a major problem in multi-center fMRI analysis. 2) Proposes a flexible harmonization model that is conditional on prompts rather than having predetermined mappings. 3) Employs a hierarchical model starting from large scales (networks) down to small scales (regions). 4) Attains state-of-the-art results, significantly outperforming other models. 5) Highlights the contribution of different components such as hierarchy, prompts, and atlas. 6) Enhances clinical diagnosis between subjects using data from different scanners. 7) Preserves individual differences while minimizing scanner effects. 8) Verifies model reliability by testing on various atlases and using qualitative assessments.
- 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) Limited novel contribution, since the framework is a combination of current paradigms (such as contrastive alignment and graph representation) but with some changes. 2) Lack of credit is given to GraphCLIP or dilated GCN-based alignment, leading to ambiguity on novelty and its connection to other literature. 3) Lack of detailed discussion on prompt design and model design, specifically on the impact of FC graphs on prompts and performance. 4) Not investigated how sensitive the model is to perturbations to prompts, which might impact the performance. 5) Little discussion on interpretability, especially in terms of neuroscientific evidence that can help interpret the results. 6) Introduction of critical terms (for instance, InfoNCE loss) without necessary explanation or citations.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript provides an innovative and solid methodology for the harmonization of multi-site and multi-atlas functional connectivity. The prompt-based approach along with the hierarchy of coarse-to-fine alignment solves some of the shortcomings of previous approaches, allowing the proposed framework to provide effective and biologically meaningful harmonization in different acquisition scenarios.
Although the architecture itself is not particularly novel and there is a need for more detail on certain parts of the methodology, such as prompt engineering and the GraphCLIP attribution mechanism, the significance of the work is evident. The research has high potential for future use in multi-site neuroscience applications and is thus suggested for acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal provides useful clarifications, but it does not sufficiently resolve the key concerns on novelty, prompt robustness, prior-work differentiation, and neuroscientific interpretability.
Author Feedback
We thank all reviewers for recognizing the novelty and practical value. We address the main concerns below. 1) R1/R2/R3: innovation, novelty, and attribution. UniFCHarm’s key innovation is target-conditioned FC harmonization: text prompts specify the target scanner/protocol and atlas-defined FC, and hierarchical graph editing performs harmonization under this target prompt. Unlike previous harmonization methods tied to a fixed site and atlas, UniFCHarm handles arbitrary site and atlas FCs. We will also clarify that GraphCLIP denotes our FC-specific graph-text alignment rather than a new contrastive paradigm, and add corresponding citations. 2) R1/R2: coarse-to-fine rationale and Yeo-7 scaffold. Coarse-to-fine harmonization is motivated by the multi-scale nature of acquisition-induced heterogeneity, which affects both subnetwork-level and ROI-to-ROI connections. A single-step correction treats these scales equally, so ROI-level changes may remove subject-specific FC patterns with site effects. UniFCHarm first harmonizes subnetwork-level connectivity and then refines residual ROI-to-ROI biases under this coarse constraint, reducing over-correction. Table 1 (A2) supports this design: removing coarse-to-fine weakens age MAE from 4.10 to 4.95.Yeo-7 is used only as a shared cross-atlas functional scaffold, mapping ROIs from different atlases to the same seven subnetworks before ROI-level refinement. We will discuss atlas-matched or learnable scaffolds as future work in the Conclusion. 3) R1/R2: data and ground-truth FC. We use 14 datasets in total: 44,493 FC-prompt pairs for model training, SRPBS traveling-subject scans for validation, and ADNI Siemens-to-GE subset for downstream diagnosis task validation. SRPBS contains 411 rs-fMRI scans from 9 healthy males across 3 scanners. The ground-truth target FC is the same subject’s FC under the target scanner, protocol, and atlas. ADNI diagnosis task validation uses BrainGNN as classifier with 360 Siemens training FCs (NC/MCI/AD = 120/120/120) and 150 GE testing FCs (50/50/50). 4) R2/R3: loss formulation and encoder details. GraphCLIP employs symmetric InfoNCE over normalized graph/text embeddings, averaging graph-to-prompt and prompt-to-graph losses. We will clarify this loss in the revised manuscript. The notation used in the paper will be clarified as: Na is the number of ROI in atlas a, Nc=7 is the Yeo subnetwork number, d is the input feature dimension, k is the latent dimension, and L is the prompt token length. In Stage 1, the text encoder maps structured prompts to prompt embeddings, and a three-layer GCN maps FC graphs to graph embeddings. In Stage 2, the Resampler and dilated GCN handle atlas-dependent ROI dimensions for ROI-level harmonization. 5) R2/R3: training details and prompt design. We will expand Sec. 3.1 with GRETNA preprocessing, ROI extraction, Pearson FC construction, and training details. λ=0.2 is selected from {0.1,0.2,…,1.0} using validation Corr/MAE. Training takes 6.4/6.1 days for Stage 1/2 on one A100.Prompts include scanner identity, TR, TE, TI, flip angle, voxel size, and FC atlas. A1 confirms prompt design: replacing prompts with scanner-specific one-hot embeddings drops ACC from 91.2 to 72.4 and degrades Corr/MAE from 0.82/0.03 to 0.78/0.04.6) R2: baseline selection. Due to space limits, the original submission reports representative baselines. Each baseline is trained separately for each source-target harmonization task under the same setting. We further include the previously obtained DeepCombat comparison (Corr 0.78, MAE 0.04, target ACC 50.20%, age MAE 4.70), indicating that nonlinear fitting helps but still lacks prompt-conditioned multi-target control. 7) R3: interpretability. Due to space limits, the original submission only briefly discussed interpretability through Fig. 3(A). UniFCHarm yields target-consistent connectivity patterns with localized and moderate modifications, indicating controlled harmonization without widespread distortions.
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 addresses an important problem in multi-site fMRI harmonization and proposes a promising prompt-guided, coarse-to-fine framework. one reviewer strongly supports acceptance. However, the current reviews also raise important concerns about insufficient technical details, unclear motivation for some design choices, limited explanation of prompt/GraphCLIP components, missing dataset and preprocessing information, and incomplete justification of baseline selection. Therefore, I recommend inviting the authors for rebuttal to clarify these key issues.
- 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.
After carefully reviewing the authors’ rebuttal, I believe that the authors have effectively addressed the concerns raised by all reviewers. The responses are clear and sufficiently detailed, and the revisions have improved the manuscript accordingly. 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.
Reviewers agreed this addresses a real and important problem, multi-site, multi-atlas fMRI harmonization, with a useful prompt-conditioned coarse-to-fine design, and the validation is strong (R1). The main disagreement between the reviewers were missing technical and experimental detail and the novelty/attribution of the GraphCLIP component (R2, R3). I believe the rebuttal handled these well, supplying the dataset and preprocessing specifics, the InfoNCE loss, the notation, the hyperparameter selection, and clearer positioning etc, and R2 moved to accept on that basis. R3 stayed at reject, but largely on points the rebuttal at least partially answered, and I noticed R3’s own pre-rebuttal write-up had actually acceptance, so I weight that final vote less.
On balance, with R1 clear accept, R2 persuaded to accept, and the substantive concerns resolved through clarification, I recommend acceptance. Please include the promised details in the camera-ready.
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.
While reviewers remain split on acceptance post-rebuttal, I believe this paper merits acceptance. The authors have addressed the reviewers’ concerns to my satisfaction. I do not believe that novelty can only come from a completely new architecture or design, and unique interpretations or applications of existing designs to overcome previous limitations are meritorious. Further, the reviewers agree that this work has substantial promise in the field.
