List of Papers Browse by Subject Areas Author List
Abstract
Segmenting acute and hyper-acute brain lesions in neonatal hypoxic ischemic encephalopathy (HIE) from diffusion-weighted MRI (DWI) is critical for prognosis and treatment planning but remains challenging due to severe class imbalance and lesion variability. We propose a computationally efficient 2D segmentation framework leveraging ADC and ZADC maps as a three-channel input to UNet++ with an Inception-v4 encoder and scSE attention for enhanced spatial-channel recalibration. To address class critical imbalance and lack of volumetric context in 2D methods, we introduce a novel boundary-and-region-aware weighted loss integrating Tversky, Log-Hausdorff, and Focal losses. Our method surpasses state-of-the-art 2D approaches and achieves competitive performance against computationally intensive 3D architectures, securing a DSC of 0.6060, MASD of 2.6484, and NSD of 0.7477. These results establish a new benchmark for neonatal HIE lesion segmentation, demonstrating superior detection of both acute and hyper-acute lesions while mitigating the challenge of loss collapse. The code is available at https://github.com/BONBID-HIE/Neonatal-HIE-SPARSeg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4455_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/BONBID-HIE/Neonatal-HIE-SPARSeg
Link to the Dataset(s)
https://zenodo.org/records/10602767
BibTex
@InProceedings{RaoAmo_Spatial_MICCAI2025,
author = { Rao, Amog AND Shukla, Ananya AND Bhargava, Jia AND Ou, Yangming AND Bao, Rina},
title = { { Spatial Prior-Guided Boundary and Region-Aware 2D Lesion Segmentation in Neonatal Hypoxic Ischemic Encephalopathy } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
page = {520 -- 530}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a computationally efficient two‑dimensional lesion segmentation framework for neonatal hypoxic‑ischemic encephalopathy from diffusion‑weighted MRI, that integrates spatial priors from ADC, ZADC, and thresholded ZADC channels into a UNet++ architecture with Inception‑v4 encoder and scSE attention. A novel boundary‑and‑region‑aware TLHF loss (Tversky, Log‑Hausdorff, Focal) mitigates class imbalance and achieves benchmark performance competitive with 3D methods, yielding superior boundary delineation.
- 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.
- TLHF boundary‑and‑region‑aware loss uniquely stabilizes training on severe imbalance, enhancing hyper‑acute neonatal lesion detection accuracy.
- Three‑channel input representation innovatively encodes spatial anatomical diffusion priors, improving lesion detectability.
- Efficient 2D UNet++ with Inception‑v4 and scSE attention matches 3D architectures’ performance, lowering computational overhead.
- 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.
- TLHF loss weight selection lacks systematic sensitivity analysis across hyperparameter settings, risking unstable segmentation performance
- Fixed ZADC<−2 threshold may misclassify borderline diffusion deviations, reducing sensitivity for early hyperacute lesion detection.
- Missing Ablation studies versus other newer 2D models like SegFormer, DeepLabv3+.
- 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 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
As mentioned in the above weakness section. It is using older methods. Also can use different fusion techniques to improvise the segmentation.
- 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.
I think the authors have justified the weakness good enough and I am satisfied with their response. However, I still believe ensembling among new models or new models with better hyperparameter tuning would outperform the existing scores.
Review #2
- Please describe the contribution of the paper
The paper introduces a segmentation framework that integrates several advanced components: UNet++ architecture, Inceptionv4 encoder, and scSE attention mechanisms. While each of these components has been individually applied in medical image segmentation, their specific combination in this work appears to be novel.
- 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.
- Computational Efficiency: The method is 2D-based, making it much more resource-efficient compared to memory-heavy 3D models, while still achieving comparable or better results.
- Novel Input Representation: Stacking of ADC, ZADC, and thresholded ZADC<−2 channels provides a richer, more lesion-focused input. The thresholded map acts as a spatial prior, boosting detection of hyper-acute lesions.
- Hybrid Loss Function (TLHF Loss): Combines Tversky, Log-Hausdorff, and Focal losses to handle class imbalance, boundary precision, and loss collapse. This is especially effective for small lesions that are often missed.
- Attention Mechanisms (scSE Blocks): The use of spatial and channel-wise attention improves feature recalibration, especially beneficial for detecting fine lesion structures.
- Benchmark Performance: Outperforms or matches state-of-the-art 3D methods on key metrics (DSC, MASD, NSD), particularly impressive given the lower computational burden.
- Robust Preprocessing and Augmentation: Effective normalization, resampling, and non-aggressive augmentation strategies ensure data consistency and enhance generalization.
- 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 Volumetric Context: Being a 2D approach, it lacks the 3D spatial coherence available in 3D models, which might limit understanding of inter-slice relationships.
- Limited Generalization to Different Modalities: The method is specifically tailored for DWI-based ADC/ZADC maps. Its performance on other MRI modalities (e.g., T1, T2) or datasets with different imaging characteristics is untested.
- Complex Training Setup: While the model is efficient, the training setup includes dynamic loss weighting and custom preprocessing, which could add complexity for real-world clinical deployment or replication.
- 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 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 paper demonstrates a solid methodological contribution with thoughtful architectural design and a well-structured presentation. It holds strong potential, particularly in neonatal brain lesion segmentation.
- 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 #3
- Please describe the contribution of the paper
The paper presents a 2D segmentation framework for detecting acute and hyper-acute neonatal HIE lesions in DWI, using a three-channel input to encode region-specific diffusion abnormalities.
- 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 topic is of high relevance to both MICCAI and clinical community, and the research introduces a novel direction. The dataset size appears reasonable, and the methodology is clearly described.
- 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.
That said, several important concerns should be addressed to strengthen the manuscript:
-
While the introduction aims to provide a comprehensive background, it becomes quite lengthy. Consider tightening the text to improve focus and readability.
-
Please include a representative image showing examples of acute and hyper-acute lesions from the dataset. This would greatly aid readers’ understanding of the task and data characteristics. this would be also particularly of interest to the medical/clinical community.
-
The rationale for training the model for 75 to 100 epochs is unclear. Additionally, there is no mention of whether early stopping was used. This information is important for evaluating training stability and the potential for overfitting.
-
In Section 2 (Dataset), please specify the age range and clinical conditions of the neonates included in the study. Similarly, it is unclear to how many samples existed in each group (acute vs hyperacute - and the time-range associated to each).
-
The limitations of using a 2D approach, especially compared to 3D segmentation, are evident in the results. It would be helpful to provide a clearer explanation of how 2D slices were selected from the 3D volumes.
-
Finally, it is unclear how large the dataset became after data augmentation. Please include details on the augmented dataset size/training split to give a full picture of the training conditions.
-
- 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?
As per the comments above,
- 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.
Authors’ rebuttal comments are satisfactory.
Author Feedback
We thank the reviewers for their constructive feedback and appreciate the recognition of our novel and clinically grounded 2D framework for segmenting neonatal HIE lesions—a critical but neglected condition. Our framework combines a 3-channel spatial diffusion prior, a region and boundary-aware TLHF loss, and scSE attention, setting a strong benchmark on the BONBID-HIE dataset characterized by acute, clinically sparse multifocal lesions.
Reviewers raised helpful questions around TLHF loss sensitivity, role of the binary prior channel, completeness of architectural ablations, and justification for using 2D over 3D segmentation, all of which we address through detailed clarifications, and empirical evidence in the responses below.
TLHF Loss Stability & Sensitivity Analysis (R2) During development, we conducted a ±20% grid sweep around our final weights (γT=1.5, γLH=2, γF∈{2,4}) and observed minor performance variation (±0.29% DSC, ±0.42% MASD, ±0.37% NSD), confirming robustness to local weight changes. To further stabilize training, we applied a curriculum-style schedule for γF—starting at 2 to prioritize region-wise learning, and linearly increasing to 4 to emphasize small, hard lesions. This coarse-to-fine ramping smooths gradient flow and aligns with established practices in lesion segmentation.
2D vs 3D Spatial Coherence & Slice Selection (R3, R1) Our scans exhibit large, uneven inter-slice gaps (2–6 mm), common in real-world clinical settings, which weakens slice-to-slice correlation and causes 3D convolutions to over-smooth distant anatomy. To mitigate this, we extract 2D axial slices (Sec 3.1 Fig 1), and Table 4 shows our lightweight 2D model choice (UNet++ ~26M params) matches or outperforms several 3D baselines (3D UNet++ ~70M). We’ll highlight this rationale in both the Introduction and Sec 3.1. Where isotropic volumes exist, exploring 3D or 2.5D models are promising future directions.
ZADC<−2 as a Non-Exclusive Spatial Prior (R2) The ZADC<−2 channel is a spatial attention prior—not an exclusive decision mask; based on clinical HIE studies where ADC values 2 SD below the normative mean indicate abnormal diffusion (Bao et al.). Borderline voxels (e.g., ZADC = −1.8) remain fully visible via the primary ADC and ZADC channels (continuous diffusion sources). Empirically adding ZADC<−2 to ADC+ZADC boosts DSC by 5.98% (DenseNet‑161, Table 1), improving sensitivity without loss of true positives.
Comparison with Modern 2D Baselines (R2) In early trials, DeepLabv3+, SegFormer and TransUNet showed relatively lower generalization on our critically‑imbalanced data (74/133 scans with <1% lesion volume). DeepLabv3+ achieved a DSC of 0.291 (ADC), 0.527 (ADC, ZADC), and 0.559 (ADC, ZADC, ZADC<−2), while SegFormer achieved 0.289, 0.531, and 0.564 respectively. Given space constraints, we focused on providing results for UNet++ (Table 1) that consistently converged and demonstrated stable performance across lesion sizes. We will clarify this in Sec 4.2 and if permitted, release full ablation results in our GitHub, as supplementary material.
- Dataset & Training (R1): As per Bao et al. (Sci Data 2025), scans are term‑born HIE neonates (mean age 3.9±2.7 days). Lesion volumes stratified into <1% (N=74), 1–5% (N=26), >5% (N=33). Early stopping patience = 15–20 epochs; convergence 75–100 epochs; on‑the‑fly non-aggressive augmentations; no change to stored dataset. Train:Test=3314:1752 2D slices.
- Representative Images (R1): Will include images for each lesion volume class for better illustration.
- Fusion Strategy (R2): We use early fusion 3-channel + scSE blocks for spatial/channel re-calibration. Mid‑level or modality‑aware fusion is promising future work.
- Generalization (R3): While focused on DWI here, our code supports arbitrary MRI contrasts; we plan to evaluate T1 and multi‑modal data in follow‑up studies.
Upon acceptance we will publicly release the code and pre-trained models to ensure complete reproducibility.
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”.
While two of the reviewers have fairly positive comments, I think reviewer #2 has raised important concerns that I would like the authors to address. Furthermore, after reading the paper myself, I have further concerns that I would like to see discussed by the authors.
First of all, the way the results are presented is confusing and fairly limited. Why is table 2 only focusing on two of the losses? Why is only one of the rows an ablation on attention? There are multiple choices and configurations possible (3 losses, and technically 3 possible attention configurations). Why is table 3 comparing attention vs not using it? I think showing results by lesion volume on the other tables would also be an interesting idea. And finally, why is the best result of table 1 not the best overall result shown in tables 2 and 4? I do understand that the authors are trying to follow some sort of incremental approach to the results, but the presentation is confusing as it looks like the choices are arbitrary.
Similarly, I also find the choices of loss function and architectural choices arbitrary and not well justified. Especially, when no baseline network is actually provided to better understand why these choices were made.
With all these concerns in mind, and the reviewer’s feedback I would like to invite the authors for rebuttal to better clarify their proposal.
- 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’
After the rebuttal, all the reviewers suggested acceptance of this work. However, there are some minor points, and the authors are expected to clarify them in the final version.
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’
Reviewers agreed on the strengths of the paper, and most of the reviewer’s concern were addressed in the rebuttal. I recommend acceptance.