List of Papers Browse by Subject Areas Author List
Abstract
The integration of neural networks with Magnetic Resonance Imaging (MRI) data for brain disease diagnosis has become a significant research focus. However, the inherent complexity of 3D MRI data poses challenges for traditional models like CNNs and Transformers, leading to high computational costs and difficulties in clinical deployment. Spiking Neural Networks (SNNs), inspired by biological neurons, offer a promising alternative with enhanced efficiency and robustness. Yet, their application to MRI data is limited by fixed time-steps that fail to account for inter-sample variability. To address this, we propose a Variable Time-Step Spiking Neural Network (VT-SNN) that dynamically adjusts the time-step based on sample-specific uncertainty. Our method employs an SNN-based Transformer module to convert MRI data into spike form and extract features, followed by a variable time-step module that measures decision uncertainty using Fisher information and PAC-Bayes theory. Experiments on AHNU and AMRD datasets demonstrate superior classification performance and reduced computational costs. Our codes are available at https://github.com/UAIBC-Brain/MICCAI-2025-Paper-VT-SNN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1878_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/UAIBC-Brain/MICCAI-2025-Paper-VT-SNN
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RaoHao_VTSNN_MICCAI2025,
author = { Rao, Haonan and Wei, Shaolong and Jiang, Shu and Wang, Mingliang and Ding, Weiping and Huang, Jiashuang},
title = { { VT-SNN: Variable Time-step Spiking Neural Network Based on Uncertainty Measure and Its Application in Brain Disease Diagnosis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {677 -- 686}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors introduce a spiking neural network Transformer designed for MRI image classification. Rather than relying on repeated Transformer blocks, they propose a streamlined architecture featuring a variable time step module to reduce complexity.
Experiments conducted on two large-scale MRI datasets demonstrate that the proposed model outperforms two convolutional neural networks and three existing spiking neural network approaches.
- 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 methodology is compelling, with a well-executed comparison to existing approaches.
- 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 lacks clarity, with several important details missing—most notably, the image encoding strategy is not specified. Additionally, the absence of source code makes the work difficult to reproduce. Although the authors attempt to explain the mathematical foundations of their method, the presentation is convoluted and ultimately difficult to follow.
The baseline artificial neural networks used in the paper appear outdated. They are limited to convolutional architectures and do not incorporate attention mechanisms, which are now standard in state-of-the-art models. A more relevant comparison would be with Transformer-based approaches, such as the one presented in this recent work on Alzheimer’s detection using MRI. https://www.nature.com/articles/s41598-024-59578-3
- 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 does not provide sufficient information for reproducibility.
- 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 the method and idea seems interesting, the presentation and lack of comparison with modern artificial neural networks guide my rating.
- Reviewer confidence
Somewhat confident (2)
- [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 #2
- Please describe the contribution of the paper
This paper presents a variable time-step spiking neural network (SNN) that allows for dynamic adjustments of which time step to terminate based on the uncertainty of the sample applied to the diagnosis of brain disease. Having a variable time-step SNN enables adaptive temporal resolution that better captures the dynamic characteristics of neural signals, allowing for allocating more computational resources to critical, fast-changing segments while reducing redundancy in slower or less informative regions.
- 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.
In the context of brain disease diagnosis, variable time-step spiking neural networks (SNNs) are a novel methodology as they closely mimic the temporal dynamics of biological neural systems. Unlike conventional neural networks and even standard fixed-step SNNs, this work introduces a variable time-step module that measures the uncertainty of the MRI samples to generate a confidence score that is then used to select the time step based on a predetermined threshold. Thus, it represents a novel and promising tool for early and accurate brain disease detection due to its adaptability between MRI samples. Additionally, the paper merges SNNs with Transformer modules, leveraging the Transformer’s global attention capabilities and the SNN’s energy efficiency. This hybrid design enhances feature extraction from 3D MRI data while maintaining economical computing, addressing a key challenge in medical imaging.
- 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.
Although the paper presented is methodologically relevant, the writing could be improved. The vast amount of equations makes the paper less readable. Some could be replaced by simpler textual descriptions or, in cases such as equation 10, which could be broken down or better described to allow an easier understanding. Figure 2 claims to highlight disease-related brain regions but lacks validation by medical experts or comparisons to established neuroanatomical findings, reducing the interpretability of the visualized “spike activations.” Some key aspects of the methodology, such as the maximum time-step and confidence score, could be further justified, as could explain the relevance of Fisher Information and PAC-Bayes theory regarding the time-step choice. Regarding the ablation study, while it is interesting to see the difference between fixed and variable time steps, the ablation could be further expanded to analyze the overall relevance of each module present in the network or critical components such as the uncertainty module, Fisher Information integration, or the SNN-Transformer architecture. This limited ablation study fails to demonstrate the effects of the various components.
- 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.
- 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?
Although the methodology brings an interesting development for the applicability of brain disease diagnosis, the results lack the necessary ablation to understand the impact of this innovation.
- 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.
Considering the rebuttal (also for the other reviewers’ comments) and the other papers I reviewed this year, I stay with my decision for a (weak) accept.
Review #3
- Please describe the contribution of the paper
summary This paper proposed a Variable Time Step Spiking Neural Network (VT-SNN) that dynamically adjusts the time-step based on sample-specific uncertainty. This method employs an SNN-based Transformer module to convert MRI data into spike form and extract features, followed by a variable time-step module that measures decision uncertainty using Fisher information and PAC-Bayes theory.
- 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 addressed the high computational costs and difficulties in clinical deployment due to the inherent complexity of 3D MRI data and its challenges for traditional models like CNNs and Transformers.
- it tried to use Spiking Neural Networks (SNNs) as an alternative with efficiency and robustness to MRI data to account for inter-sample variability.
- Experiments on AHNU and AMRD datasets demonstrate classification performance and reduced computational costs.
- 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 techniques used in this paper is well established methods, e.g. SNN, self-attention.
-
It is general to say that SNN usually does not work as Transformer based method. As shown in table 1, the proposed VT-SNN works very well. Please show detailed analysis for this issue. It would be better to release the source code for reproducibility.
- For the Time-step Selection, it set T=4. Please show analysis for this case.
- Lamda_1 and Lamda_2 are set to be larger than 0. What is the real value when it is used in this paper?
- 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 provide sufficient information for reproducibility.
- 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?
See weakness above.
- Reviewer confidence
Very confident (4)
- [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
Author Feedback
We thank all reviewers for their comments and enthusiasm for our work. Code: We will make all code publicly available. R#1 Q1:The vast amount of equations A1: Many formulas are necessary, as they are crucial for deriving the time steps that balance accuracy and efficiency for each sample. However, As suggested, we will re-decompose and describe the formulas in the final version. Q2: Spike activations A2: In Fig. 2, PD spikes mainly activate in the substantia nigra region, while SZ from MMD activates areas like the prefrontal cortex and temporal lobe, aligning with existing medical diagnoses.[Thomas R. “Rethinking schizophrenia.” Nature 468.7321 (2010): 187-193.][ Bloem B R, Okun M S, Klein C. Parkinson’s disease[J]. The Lancet, 2021, 397(10291): 2284-2303.] Q3: Confidence score and the time-step choice A3: The Fisher information quantifies uncertainty caused by differences in brain structures. Using PAC-Bayes theory, we integrate the Dirichlet loss based on the Fisher information to let the model reason with prior knowledge, avoid overconfidence, and output reliable confidence scores for variable time steps. Q4: Ablation study A4: Our work focuses on ensuring the model’s excellent performance under variable T, where the confidence scores output by the uncertainty module determine this process. Therefore, in Fig. 3’s ablation section, AMRD/AHNU (SNN) has already removed the uncertainty module, while AMRD/AHNU (VT-SNN) has significantly outperformed it in both T and ACC. Omitting detailed section(e.g., Fisher information) of ablation avoids shifting our focus to uncertainty representation capabilities, but we’ll add more experiments in future research.
R#2 Q1: The image encoding strategy A1: In Section 2.1 (method), we clearly elaborate our model by integrating formulas with the framework in Fig. 1. In the first paragraph of Section 2.1 and the SFP in Fig. 1, we illustrate the image encoding strategy: Images are first encoded by a convolutional layer and then are encoded with LIF neuron spikes for temporal coding. We will publicly release all code after the paper is accepted to ensure that anyone can reproduce our work. Q2: Baseline A2: The work from Nature you mentioned used 2D slicing of 3D MRI (over 10k slices) with ViT for excellent performance, fitting ViT’s large-data needs. Our case differs: we use 3D MRI for training. Due to limited data resources, we adopt lightweight ANN baselines to balance cost and performance for direct Power/Acc comparison. As suggested, we will add supplementary experiments in the journal version to enhance rigor.
R#3 Q1, Q2: The reasons why our work is good A1, A2: We will publicly release all code for reproducibility. Recent SNN-Transformer studies (e.g., Spike-Driven Transformer V2) match ViT’s performance. For sMRI data, brain structural variability makes fixed time steps suboptimal. Our method differs from traditional SNNs: it uses uncertainty measurement for variable time steps selection and avoids key information loss from short T and noise redundancy from long T. Our work enhances robustness and enables the first SNN-based sMRI analysis with biological interpretability. Q3: T is set to 4. A3: Typically, effective information in the brain can be transmitted via biological neurons emitting spikes in extremely short time. But in our work, if T is too small, key information will be lost, while if T is too large, noise surplus will increase. Comprehensively, T=4 can balance accuracy and efficiency simultaneously. We will add experiments in the journal version to demonstrate this point. Q4: Lamda_1 and Lamda_2 A4: Lamda_1 is 0.01 and Lamda_2 is -1. In our work, the uncertainty module determines the reliability of output confidence scores and is crucial for deciding variable time steps. By integrating Fisher information and KL divergence via PAC-Bayes theory, we use Lamda_1 and Lamda_2 to mitigate overconfidence, enhancing generalization and probabilistic interpretability.
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.
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 #3
- 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’
While the authors provided additional details in the rebuttal, one critical issue still remains. The paper lacks comparison with most recent works on this topic, especially those published in 2024. Most of the baselines selected were quite old.