List of Papers Browse by Subject Areas Author List
Abstract
Breast tumor segmentation in dynamic contrast-enhanced magnetic resonance image (DCE-MRI) achieves precise delineation of tumor boundaries and subregions by capturing rich tissue heterogeneity information. However, its reliance on contrast agents may cause adverse effects, and the complex process of acquiring complete time-series data. In contrast, current non-contrast image segmentation methods suffer from insufficient accuracy due to the lack of explicit tissue heterogeneity information. To address these limitations, we propose a approach for tumor heterogeneity estimation and segmentation in non-contrast images. The core idea is to extract tissue heterogeneity information from DCE-MRI and transfer it to a non-contrast image segmentation network, achieving tumor segmentation accuracy comparable to DCE-MRI-based methods. Our approach uses a vector quantized-variational autoencoder (VQ-VAE) based clustering model to transform images into heterogeneity maps, capturing structural features of tumor subregions. These maps serve as the ground truth for training. Then, a heterogeneity information prediction model (HIPM) estimates heterogeneity maps from non-contrast images. These features are utilized as prior information to guide the segmentation network, further improving segmentation accuracy. Experimental results demonstrate that the cluster compactness (CPN) and Davies-Bouldin index (DBN) of the clustering reach approximately 0.05 and 0.001, respectively, indicating high clustering accuracy. Our method provides intuitive visualization of tumor heterogeneity without the need for contrast agents and significantly improves segmentation accuracy, with Dice Similarity Coefficient (DSC), Positive Predictive Value (PPV), and Sensitivity (SEN) increased by 20% compared to other non-contrast image segmentation networks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1910_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/millieXie/HCNet
Link to the Dataset(s)
ISPY1 dataset: https://www.cancerimagingarchive.net/collection/ispy1
DUKE dataset: https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70226903
BibTex
@InProceedings{XieXin_Tumor_MICCAI2025,
author = { Xie, Xinyu and Han, Luyi and Li, Yonghao and Duan, Yaofei and Sun, Yue and He, Muzhen and Tan, Tao and Shen, Dinggang},
title = { { Tumor Segmentation with Heterogeneity Clustering in Non-contrast Breast MRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15961},
month = {September},
page = {642 -- 652}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a novel three-stage approach to segment breast tumors using only non-contrast T1 MRI while leveraging heterogeneity information implicitly learned from DCE-MRI during training. It first uses a VQ-VAE to cluster DCE-MRI time-intensity curves (TICs) into heterogeneity maps. Then, it trains a network (HIPM) to predict these maps from non-contrast T1 images. Finally, a segmentation network (KIS-Net) fuses features from the non-contrast T1 image and the predicted heterogeneity map to produce the final segmentation, aiming for DCE-level accuracy without requiring contrast agents at test time.
- 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 work addresses the important challenge of improving non-contrast MRI segmentation by incorporating tumor heterogeneity information, potentially reducing the need for contrast agents. I feel that the core idea of extracting heterogeneity patterns from DCE-MRI (via TIC clustering) and training a network to predict these patterns from non-contrast images for segmentation guidance is innovative. The authors state that their method demonstrates substantial performance gains compared to methods using only non contrast images. The intermediate prediction of heterogeneity maps from non-contrast data could offer additional clinical insights.
- 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 multi-stage approach proposed by the authors requires paired DCE-MRI and non-contrast MRI for training, which might limit applicability. The overall pipeline is a bit complex. The heterogeneity maps used for supervision in the prediction stage are generated via unsupervised clustering (VQ-VAE), not true biological ground truth, potentially limiting accuracy. Lack of code/data availability commitment and insufficient detail on network architectures (KIFB block), hyperparameters (loss weights), and clustering make reproduction difficult.
- 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
Clarify the architecture of the KIFB block and provide values/details for hyperparameters like β and λ. There were a few typos in the paper that I came across:
- Page 1 (Abstract): Change “propose a approach” to “propose an approach”.
- Page 2: Change “requiring not contrast agents” to “requiring no contrast agents”
- Page 2: “tumor heterogeneity in non-contrast enhanced images.” Here, “enhanced” seems contradictory/redundant; likely meant just “…in non-contrast images.”
- Page 2: Correct the referencing style “[8, 24, 25], [21]”
- Page 6: “…adopt the 3DUNext model as the baseline…” 3DUNext is incorrect here. Please go through the paper thoroughly for any other typos and grammatical errors.
- 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 paper proposes an interesting and clinically motivated idea for transferring knowledge from DCE to non-contrast MRI segmentation. The results show a clear benefit over purely non-contrast approaches. However, the method’s complexity, reliance on unsupervised clustering for intermediate supervision, potential dataset issues, lack of convincing quantitative comparison to established DCE-based methods, and significant reproducibility concerns weigh against it. While the direction is promising, the current execution and validation are not sufficiently rigorous for acceptance without substantial clarification and potentially further experiments (like ablation studies and direct DCE comparison).
- 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
Review #2
- Please describe the contribution of the paper
This paper addresses the segmentation of breast cancer tissue using DCE-MRI by leveraging a deep learning architecture that operates solely on T1-weighted breast images, without the need for contrast agents. The proposed method demonstrates both stability and accuracy in delineating cancer regions. The experimental comparisons are well-designed, and the rationale behind the hyperparameter settings is clearly presented, making the paper appear methodologically sound and logically constructed.
- 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.
Simplicity and interpretability: The core mechanism of the proposed network is simple yet effective. By leveraging Time-Intensity Curve (TIC)-based clustering, the method allows intuitive visualization of tumor-suspected regions from a given T1-weighted breast image, even without contrast enhancement.
Clinical applicability: Due to its lightweight and modality-independent design, the model demonstrates strong potential for clinical adoption. It operates solely on standard T1-weighted images, without requiring additional imaging sequences, which enhances its practicality and integration into routine workflows.
- 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.
Weakness – Reliance on TIC without accounting for inter-subject variation and temporal alignment A key limitation of the proposed method is its strong reliance on Time-Intensity Curves (TIC) as the primary feature for tumor region identification. However, TIC patterns can vary significantly across patients due to differences in tissue arrival time (TAT), vascular dynamics, and injection-to-scan delay. In breast DCE-MRI in particular, controlling for tissue arrival time is non-trivial and not well standardized. The current model does not appear to include any architectural component or preprocessing strategy to normalize or align TICs across voxels or patients. This lack of temporal control may undermine the consistency and generalizability of the extracted features, especially when applied to real-world clinical data with variable enhancement timing.
There are lesions that show little to no enhancement, or only subtle uptake. These may represent early-stage tumors or benign lesions with associated vascular changes. Although such lesions may not be clearly distinguishable on DCE curves or pre-contrast T1 images, they are important to identify for early diagnosis and timely intervention
- 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?
From an idea standpoint, it seems quite necessary and easily accessible; however, analysis of actual clinical data is required. The technical specialty does not appear to be significant.
- Reviewer confidence
Very confident (4)
- [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.
Loss of amplitude-related information: Global min/max normalization may suppress patient-specific TIC intensity variations, which could be critical for distinguishing subtle differences between benign and malignant lesions.
Timing matters: While the model claims to prioritize general kinetic trends over precise timing, temporal parameters such as time-to-peak or washout slope are essential diagnostic indicators in dynamic contrast-enhanced imaging. Interpolating or shifting TICs may distort these critical features.
Spatial fidelity concerns: Resampling to 1×1×1 mm³ may affect anatomical integrity, particularly for small lesions, where partial volume effects or interpolation artifacts can be non-negligible.
Review #3
- Please describe the contribution of the paper
The authors propose a tumor segmentation network for pre-contrast breast magnetic resonance imaging (MRI), utilizing time-intensity curves. The clinical motivation is to avoid the use of gadolinium-based contrast agent with known side-effects. There are three key steps:
1) Cluster time-intensity curves using vector-quantized variational autoencoders (VQ-VAEs) to generate enhancement maps.
2) Train a predictive model to estimate enhancement maps for pre-contrast MRI using the generated maps as ground truth.
3) Refine tumor segmentation by integrating the enhancement map with pre-contrast MRI features.
The authors show that their method is able to achieve segmentation metrics comparable to various models using pre-contrast and/or post-contrast and/or subtraction sequences. The method is validated on two publicly available datasets as well as one in-house 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.
This work blends clinical utility and methodological novelty with ease. Summarizing the strengths of the paper:
1) Interesting clinical use-case, the work proposes a solution to avoid contrast agents while retaining its two important benefits, (a) getting accurate tumor localization and (b) kinetics curve indicating malignancy.
2) Improvement over existing segmentation networks validated with various combinations of MRI sequences.
3) Novel methodology of using predicted tumor kinetics to inform hard-to-see tumors on pre-contrast MRI.
4) Clear writing with dense-information about methodology and experimentation.
- 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.
I believe that the manuscript can benefit from the following additions:
1) Posing kinetics estimation to follow standard definitions (https://radiopaedia.org/articles/breast-mri-enhancement-curves), to provide evidence of usability in clinical practice.
2) Provide information on preprocessing to standardize data, if any, as the manuscript uses cases scanner with a wide number of scanners and protocols, which has potential to affect the performance unless explicitly handled.
Not necessarily a weakness but the methodology is limited in scope of application to breast MRI.
- 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.
(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 work handles an interesting use-case, the experiments are well defined and properly validated. The authors do not commit to explicitly to releasing their code, however the paper details all the relevant information.
- 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.
Thanks to the authors for addressing the comments. I maintain my initial opinion on the manuscript and recommend acceptance.
Author Feedback
We thank all reviewers (R) for their valuable feedback and have responded with the necessary information. Q1: Code (R1, R2, R3) We have used public datasets in our experiments and will release all code on GitHub to support transparency and clinical applications. Q2(R1, R2): Data Preprocessing All images were resampled to 1x1x1 mm³ spacing. Each patient’s dynamic sequence was normalized using global min/max values from their P0 (pre-contrast) image. This way preserves crucial TIC trends (even with some values > 1) and critically aligns with test-time P0 availability. Q3(R1): Temporal Alignment We referenced absolute time, applying small shifts to interpolated TICs for robust alignment. Resultant slight TAT variations are tolerated, as our model prioritizes broader kinetic trends over precise timing. Preprocessing and consistent 1-2 min acquisition intervals in our multi-center data support generalizability. Q4(R1): Lesions The model was trained on diverse inputs: highly enhancing tumors, low-enhancing lesions, normal tissues, and peritumoral regions. It enables learning varied enhancement patterns to identify subtle lesions and normal tissue. We will further emphasize this. Q5(R2): Standard Definitions. We will add more descriptions to enhance clinical applicability. Q6(R2): Application. Breast cancer exhibits higher heterogeneity than other tumors, motivating our focus on breast MRI. Our method is applicable to other organs with hemodynamic imaging and can be extended to additional cancers. Q7(R3): Paired Data The required non-contrast MRI is the pre-contrast scan in standard DCE protocols, naturally paired and from public datasets, thus not limiting applicability. Using paired data to boost accuracy aligns with recent research trends (e.g., PLHN-TMI’24, Zhang-Cell’23). Q8(R3): Pipeline Our multi-stage design does not pursue complexity; instead, it is modular and purposeful. Each module has a clear division of labor and operates synergistically. Through deliberate decomposition (not mere stacking), it systematically addresses key challenges in pre-contrast MRI (e.g., missing hemodynamics, segmentation difficulties). This adds only 0.2M parameters over the baseline (3DUX-Net). The revision will detail its efficiency. Q9(R3): Heterogeneity Maps Voxel-wise biological ground truth for heterogeneity is inherently unattainable with MRI. While some MRI heterogeneity research exists for organs (Pancreas/brain), studies on breast tumors are notably scarce. Our method addresses this by creating maps from TICs. These maps, used as supervisory signals, offer visual aids and significantly boost non-contrast segmentation performance. This practical approach leverages available information where perfect ground truth cannot exist, thus constituting an advantage, not a limitation. Q10(R3): Comparison to DCE-based Methods We clarify ‘post-contrast MRI’ denotes DCE-MRI for baseline comparisons. Table 1 directly compares our method (inference on non-contrast MRI) against these established DCE-MRI methods. Our approach achieves comparable performance to DCE-MRI methods (e.g., PLHN-TMI’24, Zhang-Cell’23) (p>0.05) and significantly (>20%, p<0.05) outperforms purely pre-contrast methods. Q11(R3): KIFB Structure The estimated maps from pre-contrast inputs are first projected via a linear layer with a sigmoid activation, then fused with the coarse segmentation. This fusion is further refined using 3×3×3 conv blocks followed by batch normalization and a final sigmoid activation to produce the enhanced segmentation output. (in Fig.1(d) and Page 5) Q12 (R3): Hyperparameters Beta = 0.25 and lambda = 5 are detailed on page 6, and we will revise all typos to improve clarity. Q13 (R3): Clustering network Our process: 1) Encoded TICs into feature vectors with FC layers. 2) Discretized these vectors into clusters using vector quantization. 3) Reconstructed TICs from the quantized features via an FC-based decoder. (in Section 2.1)
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 paper proposes a clinically motivated and technically sound approach for tumor segmentation in pre-contrast breast MRI, aiming to eliminate the need for gadolinium-based contrast agents. The method introduces a simple yet effective framework that leverages time-intensity curve (TIC) clustering via vector-quantized variational autoencoders (VQ-VAEs) to generate enhancement maps. These are used to train a predictive model and refine segmentation using only pre-contrast T1-weighted images.
The approach achieves segmentation performance comparable to models that rely on post-contrast or subtraction sequences, and it is validated across two public datasets and one in-house dataset, supporting its generalizability. Its simplicity and interpretability—particularly the intuitive visualization of tumor-suspected regions—make it especially appealing. Furthermore, the model’s lightweight, modality-independent design enhances its practicality and potential for clinical adoption, as it integrates easily into existing workflows without requiring additional imaging sequences.
Overall, this is a timely and impactful contribution with clear clinical relevance, solid technical execution, and strong potential for real-world use. I recommend acceptance.
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’
There are clear concerns regarding aspects such as normalization and normalization which remain unaddressed in the response.