List of Papers Browse by Subject Areas Author List
Abstract
Creating fully annotated labels for medical image segmentation is time-consuming and expensive, underscoring the need for efficient labeling schemes to alleviate the workload. Eye tracking presents a cost-effective solution, seamlessly integrating into radiologists’ workflows while offering task-relevant eye gaze supervision. However, due to the inaccuracy and ambiguity of gaze, it may introduce erroneous supervision and hinder the model’s ability to learn robust features. To address these challenges, we propose the graph-based neighbor-aware network (GNAN). The network constructs a graph structure from the image, separating different categories of nodes by simulating the attention distribution during the diagnostic process, to learn image segmentation based on the radiologist’s gaze information. The GNAN comprises neighbor-aware pseudo supervision (NAP) and graph contrastive decoupling (GCD). NAP utilizes the neighbor features of graph nodes to infer pseudo-labels for uncertain regions, effectively compensating for the inaccuracy in gaze supervision and further refining the supervisory signal. GCD decouples the graph structure by maximizing the inter-class node feature differences to distinguish between different categories, thereby improving segmentation performance. Experimental results on the public dataset demonstrate that GNAN outperforms state-of-the-art methods. Our code is available at https://github.com/IPMI-NWU/GNAN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3730_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/IPMI-NWU/GNAN
Link to the Dataset(s)
https://github.com/med-air/GazeMedSeg/tree/main/GazeMedSeg
https://github.com/med-air/GazeMedSeg?tab=readme-ov-file
BibTex
@InProceedings{WuSha_Graphbased_MICCAI2025,
author = { Wu, Shaoxuan and Chen, Jingkun and Jin, Zhuo and Zhang, Peilin and Gao, Zhizezhang and Feng, Jun and Zhang, Xiao and Shen, Dinggang},
title = { { Graph-based Neighbor-Aware Network for Gaze-Supervised Medical Image Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
page = {229 -- 239}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper describes a method for medical image segmentation based on gaze data supervision. The technique employs graph neural network blocks and uses neighbor-aware pseudo supervision and graph contrastive decoupling to try to deal with the noisy nature of the gaze data supervision. The method is evaluated on two different public segmentation datasets (polyp and prostate) and compared with other fully-supervised and weakly-supervised 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 method is novel as far as I know The paper was generally well-written The results show improvements over other weakly-supervised approaches Evaluation was performed on two different segmentation tasks
- 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.
There was a lack of clarity about hyperparameter optimisation
- 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.
- 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
Although an anonymised link to source code has been provided, I think that the description in the paper of the hyperparameter optimisation process is limited. See below for detailed comments.
- 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?
The paper is well-written and the method appears to have some novelty, although I am not an expert on gaze-supervised segmentation methods. The results seem to be good and show improvements over other weakly-supervised approaches including SOTA gaze supervision techniques.
Overall, I think this is a reasonable paper which makes a useful contribution to the field of weakly-supervised segmentation. I do have a few concerns though, which are detailed below.
Section 3.1 lists a number of hyperparameter values for the proposed technique. How were these values chosen? I.e. which optimisation strategy, which ranges of values, which data were used? This is important when reporting experimental results to promote reproducibility. Similarly, what hyperparameter optimisation strategy was performed for the comparative approaches? It is important for the reader to believe that the comparative evaluation approach enables all methods to achieve their best performance and hence the differences observed are real and not just due to suboptimal hyperparameters.
Section 2.1 describes the GNAN architecture. But one thing was not clear to me. Is just x_s used as input to the model? The text suggests this but Fig 1 suggests that both x and x_s are used as inputs. Also, what is the motivation for using x_s as input, either on its own or in combination with x? I didn’t see any ablation test for measuring the impact of this approach and it was not justified or motivated as far as I could see. Therefore, this seemed quite an arbitrary thing to do. Please add some justification for this component of the method.
Finally, Table 1 reported the results of the comparative evaluation which looked good. But this table would be more convincing if statistical tests were included to demonstrate that the differences seen are significant.
- 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 #2
- Please describe the contribution of the paper
Creating fully annotated labels for medical image segmentation is both time-consuming and expensive, highlighting the need for more efficient labeling methods. This work introduces the graph-based neighbor-aware network (GNAN), which leverages cost-effective eye tracking data by constructing a graph from the image and employing two modules: neighbor-aware pseudo supervision and graph contrastive decoupling to refine supervision from noisy gaze data and enhance segmentation performance.
- 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 GNAN’s combination of graph-based representations with specialized modules for pseudo supervision and contrastive decoupling appears to be a good contribution.
-
The paper is well written and well organized with sufficient experimental information.
-
- 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) Unclear Motivation for the Graph-Based Approach: The paper does not clearly justify why a graph-based model is favored over traditional segmentation models (for example, transformer-based segmentation). It remains unclear whether the graph-based approach offers a significant advantage or if similar results could be achieved using more conventional methods.
2) Limited Explanation of Technical Novelty: Many of the model’s components—such as the use of GNN, graph contrastive decoupling, and leveraging human attention—are derived from existing methods. The manuscript would benefit from a more detailed discussion on what makes the approach technically novel beyond these established techniques.
3) Missing Technical Details on the Graph Structure and Computational Complexity: The paper lacks sufficient details on the graph structure. For instance, it does not describe the impact of incorporating additional nodes and weights, nor does it address how graph processing and contrastive decoupling might increase training complexity and inference time—especially in resource-limited environments.
4) Other evaluation metrics: Have the authors used other metrics for evaluation of segmentation performance such as IoU, Average Symmetric Surface Distance, etc.?
5) λ Parameter in Equation 10: The role and significance of the λ parameter in Equation 10 are not explained.
- 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.
- 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?
Overall the paper is well written and organized. The proposed method is sound and has a significant improvement over the compared models. Thus, the work is acceptable. However, there is scope for further clarification (see weaknesses).
- 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 #3
- Please describe the contribution of the paper
The paper proposed leveraging eye gaze as a supervision signal for training a segmentation network. The gaze attention maps are divided into certain (foreground and background regions) and uncertain regions. The images are represented as graphs, with nodes representing image patches and KNN used to establish graph connectivity. They adopt partial (specific to certain regions only) cross entropy as a segmentation loss. The propose Neighbor-Aware Pseudo Supervision (NAP) that utilizes the neighbor features of graph nodes to infer pseudo-labels for uncertain regions; the more similar the nodes are, the more they inherit the labels of stable nodes (those similar to nodes of certain 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.
- They test their method on polyp segmentation from gastrointestinal images and prostate in MRI.
- They compare their results (in terms of accuracy and annotation time) to several methods that use Gaze and other supervision (bounding box, points, scribbles). They show that the proposed method outperforms these competing methods by around 3-7% points, while reducing annotation time compared to full supervision and compared to other annotation types (e.g. less time than scribbles, points, and bounding boxes).
- At a high Level, the paper is well organized tables and figures.
- The method is tested on two datasets and shown superior to several competing methods.
- Ablation study shows that their contributions (UC,NAP,GCD) were effective.
- 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 clarity in describing some technical details.
- 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
It wasn’t clear to when this training will occur. There were ambitious parts in describing the method. Also, the high level setup of how and when gaze is collected and how it relates to the training was not clearly presented.
How is the feature map divided into nodes? It seems based on a rectilinear tiling. But please state that explicitly.
Are the thresholds Tf and Tb applied to the gaze attention map ga?
In equation 4 (also in the paragraph above eq. 1), clarify how is a node (e.g., v_1, v_u) encoded? Is it a spatial location (x,y,z), a feature vector (how is it obtained), or something else? I assume it is a feature vector but I don’s see this explicitly stated.
I am seeing a contradiction on how Eq1 performs aggregation and transformation on the nodes, whereas eqn 8 encourages the nodes (before and after the aggregation and transformation) to have similar features. Please clarify.
Is the step of gazing/looking at the images and collecting and analyzing the gaze maps a preprocessing step that takes place and is completed before training the segmentation network (to minimize the losses in eqn 10)?
How were the hyper parameters selected (thresholds, temperature, lambdas)?
What are the symbols on the left of table 1 (triangle, clubs, diamonds, spades, squares). I don’t see them mentioned or used anywhere else.
- 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?
The only important problem is lack of clarity of some details.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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”.
After careful consideration of this manuscript along with the expert reviews from the double-blind review process, I find the paper’s contributions to be potentially valuable to our field. While the reviewers have raised some valid points that need attention, I believe these can be adequately addressed. I recommend acceptance of this submission with the authors implementing the necessary revisions based on the reviewers’ feedback to strengthen the final version of the paper.