List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of Pelvic Radiation Injury (PRI) from Magnetic Resonance Images (MRI) is crucial for more precise prognosis assessment and the development of personalized treatment plans. However, automated segmentation remains challenging due to factors such as complex organ morphologies and confusing context. To address these challenges, we propose a novel Pattern Divide-and-Conquer Network (PDC-Net) for PRI segmentation. The core idea is to use different network modules to “divide” various local and global patterns and, through flexible feature selection, to “conquer” the Regions of Interest (ROI) during the decoding phase. Specifically, considering that our ROI often manifests as strip-like or circular-like structures in MR slices, we introduce a Multi-Direction Aggregation (MDA) module. This module enhances the model’s ability to fit the shape of the organ by applying strip convolutions in four distinct directions. Additionally, to mitigate the challenge of confusing context, we propose a Memory-Guided Context (MGC) module. This module explicitly maintains a memory parameter to track cross-image patterns at the dataset level, thereby enhancing the distinction between global patterns associated with the positive and negative classes. Finally, we design an Adaptive Fusion Decoder (AFD) that dynamically selects features from different patterns based on the Mixture-of-Experts (MoE) framework, ultimately generating the final segmentation results. We evaluate our method on the first large-scale pelvic radiation injury dataset, and the results demonstrate the superiority of our PDC-Net over existing approaches.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5402_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{XioXin_PDCNet_MICCAI2025,
author = { Xiong, Xinyu and Cao, Wuteng and Wu, Zihuang and Zhang, Lei and Gao, Chong and Li, Guanbin and Qin, Qiyuan},
title = { { PDC-Net: Pattern Divide-and-Conquer Network for Pelvic Radiation Injury Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a novel framework (called PDC-Net), which is designed specifically for Pelvic Radiation Injury (PRI) segmentation from MRI scans. It addresses challenges like complex organ morphologies and confusing context through innovative modules like Multi-Direction Aggregation (MDA), Memory-Guided Context (MGC), and an Adaptive Fusion Decoder (AFD).
- 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 proposes several novel modules (MDA, MGC, AFD) tailored to handle PRI segmentation challenges effectively.
- Extensive experiments on a large-scale in-house dataset show that the proposed work is superior over existing 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 rationale behind the effectiveness of the proposed Memory-Guided Context (MGC) module in reducing interference from irrelevant background information is unclear. A more detailed theoretical or empirical explanation is needed.
- The statement, “We then concatenate these expert features along the channel dimension and use a 1×1 depthwise convolution to further eliminate redundant features,” lacks justification. It is not evident how this operation achieves redundancy reduction—clarification or supporting evidence would strengthen this claim.
- The assertion that the feature f_patch^w is organ-related requires more support. It is recommended to include visualizations or qualitative analyses to substantiate this point.
- The explanation that “this time dimension can also be viewed as the dataset dimension” is ambiguous.
- The initialization for the memory bank M is not described. Clarifying how this component is initialized and updated during training would improve the reproducibility and transparency of the method.
- The acronym “DW” is used repeatedly (e.g., in DWConv) but is not defined. It should be clearly stated that it refers to “depthwise” convolution.
- The reported results appear to be based on a single training-testing split. To better evaluate the generalizability and robustness of the proposed method, it is strongly recommended to perform cross-validation. This would provide a more reliable assessment of model performance.
- 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?
- Several important methodological details are missing and require clarification (as outlined above).
- For the experiment part, it is recommended to perform cross-validation to verify the robustness and generalizability of the proposed approach.
- 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 addressed most concerns, so I recommend that this work can be acceptable.
But some statements, e.g., “We then concatenate these expert features along the channel dimension and use a 1×1 depthwise convolution to further eliminate redundant features”, still lack justification. And it is strongly recommended to perform cross-validation to validate the method’s robustness.
Review #2
- Please describe the contribution of the paper
For the first time, the paper proposed an automatic segmentation network for pelvic radiation injury (PRI) based on an AI model. To solve the problems existing in the PRI segmentation task, such as diverse organ structures (strip or circular structures), mixed foreground and background, and different shapes and sizes of various organs, three modules were designed: multi-directional aggregation module (MDA), memory-guided context (MGC), and adaptive fusion decoder (AFD). Specifically, MDA enhances the modeling ability of target shape through strip convolution in four directions, MGC maintains a memory parameter to store the global information of the entire data set, enhances the distinction between positive and negative categories, and reduces the interference of confusing background, while AFD constructs multiple types of decoders to realize dynamic selection and interaction of multi-stage features.
- 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 research objectives of the paper are clear, and the network architecture design is targeted and innovative. In view of the three major challenges in the automatic segmentation of pelvic radiation injuries, the corresponding network structure design is proposed: through the multi-directional strip convolution module, the problem of organ structure diversity is effectively dealt with; the global memory mechanism is introduced to alleviate the category confusion phenomenon; at the same time, an adaptive fusion decoder is designed to improve the modeling effect of complex organ morphology.
- Sufficient ablation experiments: For the three proposed modules, a comparative experiment was carried out by module replacement, and the independent contribution of each module was verified through nine groups of experimental results.
- Clinical value: The automation of PRI segmentation tasks can assist doctors to reduce manual labeling errors and improve the efficiency of injury assessment after radiotherapy, and has clear clinical application potential.
- 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 dataset is not public. The experiment is based only on an internal dataset, and the generalization is not verified on a public dataset, which may affect the objectivity and fairness of the results.
- The explanation of the network structure is limited. How is the “memory parameter” of the MGC module initialized? How is it updated? Is it learnable? In addition, MGC is essentially a regional weighted selector. Why is it only used in the last layer? Further explanation is needed.
- The quantitative comparison experiment is not sufficient. The FLOPs and the number of parameters of the comparison model are not given, and it is impossible to determine whether the Patch Shuffle operation of AFD is more time-consuming.
- 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.
(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 dataset is not public. The experiment is based only on an internal dataset, and the generalization is not verified on a public dataset, which may affect the objectivity and fairness of the results. The above significant weakness results in a final overall score of ‘Weak Accept’.
- 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.
Q6 lists the major strengths. The author has fully explained concerns in the rebuttal.
Review #3
- Please describe the contribution of the paper
-
The authors introduce the first large-scale dataset specifically curated for pelvic radiation injury segmentation, addressing a critical gap in the medical imaging domain and enabling further research in this area.
-
Problem Decomposition into Three Key Components The segmentation task is systematically divided into three complementary modules to better capture multi-level contextual and semantic information:
a) Feature Aggregation Module: Integrates multi-scale features from different backbone stages to enhance local-global representation.
b) Dataset-Level Memory-Guided Context Module: Introduces a memory mechanism that stores dataset-level contextual priors to improve robustness and consistency, particularly in challenging regions.
c) Mixture-of-Experts Architecture: Employs expert-customized depthwise convolutions with varying kernel sizes to specialize in different spatial patterns, effectively fusing directional and multi-scale context for accurate 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.
1) The paper introduces an innovative extension of traditional strip convolutions by incorporating diagonal filtering paths. This design improves the model’s ability to capture complex and irregular anatomical structures, which are commonly observed in pelvic radiation injury.
2) Dataset-Level Memory Module for Contextual Awareness: A key strength is the use of a memory-guided context module that operates at the dataset level rather than the frame level. This allows the model to store and retrieve prior knowledge relevant to the region of interest (ROI) across the dataset.
3) Mixture-of-Experts-Based Feature Fusion Architecture The authors design a feature fusion network that utilizes a Mixture-of-Experts (MoE) architecture, where each expert is customized with depthwise convolutions of varying kernel sizes. This enables the network to dynamically leverage different spatial receptive fields, allowing it to specialize in capturing both local details and global context.
4) First Large-Scale Pelvic Radiation Injury Dataset The creation of the first large-scale dataset tailored for pelvic radiation injury is a significant contribution to the field.
- 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 Comparison with Deformable Convolutions in the MDA Module While the proposed MDA (Multi-Directional Aggregation) module using diagonal strip convolutions is novel, it would strengthen the paper to include deformable convolutions as a baseline in the ablation study. Deformable convolutions [Dai et al., ICCV 2017] are well-known for their ability to adaptively sample irregular spatial regions, making them highly effective for segmenting deformable or irregular structures. A direct comparison would better highlight the advantages and limitations of the proposed strip-based method.
-
Insufficient Description of the Memory-Guided Context Module The dataset-level memory module is one of the central contributions of the paper; however, the construction and operational details of the memory bank are under-described. Specifically, how the memory bank is initialized and updated over time remains unclear.
-
No Discussion of Model Complexity or Parameters Although the model demonstrates competitive performance in Table 1, there is no discussion of model size or computational cost, which is important in approach comparison. It would give a clearer understanding of the trade-off between accuracy and efficiency.
-
- 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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?
The method is well-designed, and first large-scale benchmark in this domain, addressing a clear gap in pelvic radiation injury analysis. However, some parts of the methodology are unclear (see major weaknesses) and require further clarification to fully verify the reported performance.
- 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
Reproducibility
Our data and code will be available.
Method Clarity
R#ALL: Initialization/Update of the Memory Bank A: We leverage the Masked Average Pooling operation to obtain rough representations of foreground class, thereby achieving memory initialization. For memory updating, we adopt an Exponential Moving Average strategy to gradually incorporating refined features into the memory: $\hat{M}i = \alpha * M_i + (1 - \alpha) * \Phi(f^w{patch})$, where ${M}i$ denotes the i-th item in the memory, $\hat{M}_i$ denotes the updated memory, $\alpha$ denotes the momentum (set to 0.8), $\Phi(f^w{patch})$ denotes the region-of-interest features mapped by a fully connected layer. R#3: Explain “this time dimension can also be viewed as the dataset dimension” A: The meaning of this sentence is to draw an analogy between the memory in video processing tasks and our MGC. In video processing, a memory module can be designed to maintain the knowledge between video frames at different time points (time dimension). Similarly, our MGC can maintain the knowledge between different MRI slices (dataset dimension). R#3: Rationale of MGC A: The core idea of the MGC module is to introduce a learnable memory bank that dynamically accumulates and reinforces the semantic knowledge relevant to the positive/negative classes. Due to severe class imbalance, simply expanding the receptive field [15] or using attention mechanisms [16] may cause valuable foreground information to be overwhelmed by irrelevant background information. Therefore, our MGC first uses the patching operation to encourage the network to select patches related to the foreground and further dynamically updates the knowledge of these patches into the memory module, thereby maintaining dataset-level foreground class information to achieve more accurate segmentation results. R#3: Redundant Feature Elimination A: The goal of the 1×1 DWConv is to reduce the 512-channel features obtained from the concat operation to 128 channels. Following the common practice [5,10], we regard the channel reduction operation as a redundant feature elimination process. R#1: Location of MGC A: Since the goal of MGC is to maintain class-wise high-level semantic information, it is not applied to the shallow encoder, the focus of which is to recognize local patterns such as textures. R#3: Acronym “DWConv” A: DWConv denotes DepthWise Convolution.
Results Superiority/Robustness
R#1/4: FLOPs and Params A: We can’t report specific values due to the rebuttal guideline. However, our PDC-Net is more efficient than the other methods for the following reasons: 1) The strip conv design in MDA is faster than the square kernel conv in other methods. 2) The memory module in MGC allows us to simply maintain dataset-level context without the computationally intensive multi-branch, multi-scale context processing in other methods. 3) The MoE design allows us to activate only a subset of experts during inference, thereby improving efficiency. R#1: Patch Shuffle A: Generally, the latency of the Patch Shuffle operation is comparable to that of a simple 3×3 convolution, which is highly efficient. R#4: Deformable Conv (DConv) Baseline A: Our MDA is more suitable for PRI Segmentation than DConv. Specifically, DConv introduces additional learnable offsets, which not only increase computational overhead but also tend to slow down convergence. In contrast, our MDA, which benefits from anatomical priors, leverages predefined four-direction strip convolution patterns to more accurately extract organ features. R#3: Perform cross-validation A: Many works [10,12] also adopted single split verification, which is acceptable.
Future Experiments
R#1: Results on public dataset A: We are the first work on PRI segmentation, so there are no public datasets. If accepted, we will compare other tasks in an extended version. R#3: Organ-related Feature A: Thanks for the suggestion. If accepted, we will provide a visualization for f_patch^w.
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.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Please address the concerns raised by the reviewers in the camera ready.