List of Papers Browse by Subject Areas Author List
Abstract
Automatic detection of abnormal cervical cells from Thin-prep Cytologic Test (TCT) images is a critical component in the development of intelligent computer-aided diagnostic systems. However, existing algorithms typically fail to effectively model the correlations of visual features, while these spatial correlation features actually contain critical diagnostic information. Furthermore, no detection algorithm has the ability to integrate inter-correlation features of cells with intra-discriminative features of cells, lacking a fusion strategy for the end-to-end detection model. In this work, we propose a hypergraph-based cell detection network that effectively fuses different types of features, combining spatial correlation features and deep discriminative features. Specifically, we use a Multi-level Fusion Sub-network (MLF-SNet) to enhance feature extraction capabilities. Then we introduce a Cross-level Feature Fusion Strategy with Hypergraph Computation module (CLFFS-HC), to integrate mixed features. Finally, we conducted experiments on three publicly available datasets, and the results demonstrate that our method significantly improves the performance of cervical abnormal cell detection. The code is publicly available at https://github.com/ddddoreen/HyperMF2-Cell-Detection.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0793_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ddddoreen/HyperMF2-Cell-Detection
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiJin_HighPrecision_MICCAI2025,
author = { Li, Jincheng and Dong, Danyang and Zheng, Menglin and Zhang, Jingbo and Hang, Yueqin and Zhang, Lichi and Zhao, Lili},
title = { { High-Precision Mixed Feature Fusion Network Using Hypergraph Computation for Cervical Abnormal Cell Detection } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
page = {252 -- 261}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a new approach to Cervical Abnormal Cell Detection, by combining two unique contributions in their approach. Specifically, Multi-level Fusion Sub-network (MLF-SNet) and Cross-level Feature Fusion Strategy with Hypergraph Computation module (CLFFS-HC). They provide a rather detailed and easy to follow description of their methods and evaluated them on three datasets with decent comparisons.
- 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.
Clear description of the algorithms. Good illustrations. Good basic evaluation setup with different datasets and other 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 two key assumptions of their motivation to this method have no references, nor are they shown experimentally to be true. I mean this section: “1) In clinical practice, pathologists typically assess target cells by comparing them with their surrounding cells, particularly when the cells exhibit ambiguous classifications. However, existing detection algorithms generally lack effective modeling of the visual feature correlations between cells, overlooking the important diagnostic information contained in spatial correlation features. 2) Cellular lesions are often accompanied by morphological abnormalities in the nucleus, such as enlargement, irregular membranes and unclear boundaries.” This results in a weak motivation to come up with their algorithm and architectural choices.
Their related work and comparisons are limited in approach types discusses and evaluated: Comparisons are limited to a single approach type to cell detection. Nowadays, people leverage all sorts of approach to detect and classify cells. i.e. HoverNet and subsequent releases like HoverNeXT leverage instance segmentation and classification approach which would be easily compared here or approaches like StarDist or CellPose. Also, approaches based on foundation models gain traction or evaluating approach using centroid detection. While I found this paper to be released after submission deadline, this could be a good place to start to update the related work and/or compared methods: https://arxiv.org/abs/2504.06957
Metrics used are limited by class imbalances in the different datasets, that’s why usually in object detection task IoU and mAP are evaluated as well as AP and AR. When only AP and AR are used, it is common to give it for several separate IoU values in order to estimate when the model’s reliability drops off. Further, no p-values to show significance of improvements are provided.
Figure 3 makes me question the dataset choices, as the ground truth is arguably worse off than some of the compared methods. Given the small and limited sections in this figure, it is not clear if their method is actually a better model or just gives worst bounding boxes than others that are closer to the ground truth.
- 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
Figure 1:(A) why is the input data only “Synthetic Data”? Does this mean it was trained with synthetic data or the data is only synthetic for the figure?
Figure 1 it is unclear from the figure alone if the distance threshold is based on spatial or feature distance. Further, it is unclear what the threshold lambda is from the paper. Might there be an optional lambda? While this might be in the promised repo (along with an ablation on lambda), it limits the reproducibility by the paper alone, as also other variables and parameter remain underspecified.
Table 1 It is unclear why not all comparisons have values for all datasets like the SOTA for A [25]. Generally, it is also preferable to keep the real dataset names in the paper to avoid confusion and the need to look it up, which would improve readability.
- 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?
Overall, the paper present innovative idea on how to solve cell detection via a hypergraph-based network. However, the limitations in the motivation for these novelties and evaluation overweight.
- 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 that most problems I had with the paper would be resolved with the answers in the rebuttal, minor corrections to be done in writing, and making the code available for future evaluation; I tend more to accept than rejection.
Review #2
- Please describe the contribution of the paper
A method for detecting abnormal cervical cells in Thin-prep Cytologic Test images. The method adapts YOLOv11 with a hypergraph convolutional network to comparing cells against each other, improving anomaly detection.
- 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 identifies a need for improved analysis methods for Thin-prep Cytologic Test images.
The creative integration of a hypergraph convolutional neural network leads to a substantial improvement over baselines.
- 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 main problem with the paper is its writing, which makes it difficult to understand the method. In particular, Section 2.2 needs a lot of improvement. Many symbols are introduced without being defined, and there are mistakes in the Equations. Specifically:
- What are the feature points x_u?
- In Equation 2, it says if V in e, but V is the set of all vertices, so would not be in e, which is a subset of vertices. The same problem occurs in Equation 3.
- Giving the dimensionality of each symbol would make this section much easier to follow, for example, is the trainable parameter Theta_e a scalar, vector or matrix?
- The meaning of capital X with a subscript is not given.
Additional writing problems are:
- The sentence “The Euclidean Norm represents all such hyper-edges…” does not make sense. The Euclidean norm is a function, it cannot represent anything.
- Features are extracted from a backbone, but this backbone is never defined.
- It’s difficult to grasp the meaning of the terms inter-correlation and intra-discriminative. I think inter-cell correlation and intra-cell discriminative would be easier to comprehend.
- 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.
- 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
Typos: Abstract: “effectively fuse different” -> “effectively fuses different” Page 2: “cervical abnormal cells” -> “abnormal cervical cells” Page 6: “Experimental Details:Our” -> “Experimental Details: Our” Page 8: “tion.This work” -> “tion. This work”
- 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 is powerful, but the method section currently has mistakes and is difficult to follow.
- 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.
I believe the proposed method is a valuable contribution. My main complaint was with the writing of the original manuscript. The authors acknowledged my comments in the rebuttal and have promised to correct mistakes in the final version.
Regarding the term “cervical abnormal cells,” it’s obviously not a big deal, but I would still recommend writing “abnormal cervical cells,” regardless of any other paper’s choices.
Review #3
- Please describe the contribution of the paper
The authors introduce an innovative mixed feature fusion network that employs hypergraph computation for detecting cervical abnormal cells. The network utilizes a multi-level fusion subnetwork with three types of convolution to enhance the diversity of feature representation. Additionally, a cross-level feature fusion strategy, incorporating a hypergraph computation module, is used to merge inter-correlation features of individual cells with intra-discriminative features at the population level. The authors validate the effectiveness of their approach using three open 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.
- Key enhancements to yolov11, yielding major performance improvements with respect to state of the art approaches
- Description of the method is mostly exhaustive
- 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.
- It is unclear where the cell detection process occurs, and why the diagrams depict feature distances between cells instead of general image patches. Are the authors assuming that the patches represent cells, or has the detection already been performed?
- The introduction of graph construction in the method section is somewhat fragmented. Edges are introduced only later on, which makes the explanation harder to follow. The authors should consider revising the presentation for better clarity and flow.
- In Table 1, the state-of-the-art value for C - AP is incorrect and should be 35.9.
- In Figure 1, the sliding window approach is not accurately represented. It appears that patches are selected in a non-systematic manner and vary in size. A more accurate representation would involve a grid-like structure, unless the text describing the process is inconsistent with the actual method.
- 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?
Approach looks strong based on the results but exposition and figures should be revised to improve clarity
- 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
We thank all the reviewers for acknowledging the novelty, and for providing constructive comments. We summarize and respond them as follows: Q1: Position A1: Our goal is to detect lesion cells from a WSI. Clinically, an important objective of screening is to identify and classify them into different lesion types. This indicates that lesion cells can be directly detected from cervical cytology images without the need for precise cell segmentation. Additionally, compared to the instance-level bounding box annotations required by object detection methods, the pixel-level annotations needed for segmentation methods are more difficult to obtain and involve higher annotation costs.
Q2: Motivation A2: These two motivations are derived from the clinical screening experience of pathologists and evidence from medical research literature. Object detectors like YOLO11 have limitations in capturing the global correlations between cells. Graphs can only represent binary relationships between cells. However, cells in WSIs exhibit multiple types of interactions, such as overlapping cells and clustered cells, which GCN struggle to effectively model. To address this challenge, we use hypergraphs to model the high-order semantic information between cells and effectively fuse these contextual features with the discriminative features within cells. We will further clarify this point in the revised version.
Q3: Table Presentation A3: We selected three public cervical cell datasets from the past five years. Due to the limited width of the template, the names of the three datasets cannot all be listed in the same table, so we use code names for identification. Similarly, due to the limitation of width, we regret that we were unable to present the AP of the model under different IoUs. The trained model has been open-sourced at the code link provided in the original paper for validating the model’s superiority. We compared our method with classical object detectors and the existing SOTA models on each dataset. Therefore, the SOTA metrics do not represent the optimal performance of classical object detectors but rather the metrics of the SOTA models corresponding to each dataset. Thus, the latest C-AP values we provided are correct. Thank you for your suggestion, and we will optimize the chart presentation in the revised version.
Q4: Figure Presentation A4: In Dataset C, the input data includes the original images and their augmented versions using imgaug, the “synthetic data” referred to as Figure 1(A). Due to the excessively large pixel size of Dataset A, we cropped the WSI into 640×640 images using a sliding window and excluded data without labels. The preprocessed data has been open-sourced in the repository link provided in the original paper, and we will optimize the figure presentation in the revised version.
Q5: Ablation Study on lambda A5: This is a constructive suggestion. Section 2.2 illustrates that lambda is a distance threshold based on the distance of feature maps. Due to space constraints, we were unable to present the results of experiments related to lambda. In fact, the detection accuracy exhibits a convex function relationship with the setting of lambda. We have open-sourced the code to validate the superiority of the model in cervical cell detection.
Q6: Specific Wording A6: Thank you for your suggestions. We mistakenly denoted v as V in Formulas 2 and 3, and some sentences lacked spaces at the beginning. We will use more rigorous expressions in the revised version. Additionally, the “backbone” is a fixed component of the YOLO framework, specifically used for feature extraction, and we have adopted this terminology here. X represents the feature map extracted by the backbone, and X_u denotes the feature vector of a feature point u. Furthermore, we use the term “cervical abnormal cells” with reference to the paper Robust Cervical Abnormal Cell Detection via Distillation from Local-Scale Consistency Refinement [2023 MICCAI].
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’
This manuscript introduces a hypergraph-based cell detection framework, which combines spatial correlation features and discriminative features. The method produces very promising cervical cell detection performance on three public datasets. In addition, the rebuttal has addressed the reviewers’ major concerns regarding the motivation of the study and the presentation of the method and experiment results.
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’
The paper presents a novel mixed feature fusion network using hypergraph computation for cervical abnormal cell detection, enhancing YOLOv11 with Multi-level Fusion Sub-network and Cross-level Feature Fusion Strategy. Strengths include (1) innovative integration of hypergraph convolutional neural networks achieving substantial performance improvements; (2) clear algorithmic descriptions with comprehensive evaluation on three datasets; and (3) effective modeling of high-order cell relationships addressing traditional detector limitations. Weaknesses include (1) initial unclear motivational assumptions and writing clarity issues in mathematical formulations; (2) limited comparisons with recent cell detection approaches; and (3) presentation problems in figures and tables. Despite initial weak reject/weak accept ratings, all three reviewers changed to accept after the authors’ rebuttal adequately addressed major concerns and promised corrections, recognizing the method’s valuable technical contribution.