acceptodds
Withdrawn from ICLR 2027

HGMamba: A Heterogeneous Graph Mamba Framework for Interpretable Missing Modality Brain Tumor Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal MRI provides complementary information for brain tumor segmentation. In clinical practice, however, one or more modalities are often unavailable, which severely degrades models trained with complete modalities. Existing methods address this problem in limited ways. Some substitute missing modalities with zero-filled inputs, leading to biased feature representations. Others adopt multi-stage optimization with additional pretraining, increasing computational cost. Moreover, most methods overlook modality-specific characteristics and treat all modalities as equally informative for different tumor regions. To address these limitations, we propose HGMamba, an interpretable heterogeneous graph framework with Mamba-based interaction for missing-modality brain tumor segmentation. HGMamba models imaging modalities and hierarchical tumor regions, including the whole tumor, tumor core, and enhancing tumor, as heterogeneous graph nodes to capture modality-region dependencies. A data-driven affinity prior is introduced to quantify modality-region discriminability without additional learnable parameters and guide modality-to-region aggregation. A lightweight Mamba-based operator further enables efficient graph interaction, while a reliability gate and nested region interaction mechanism reduce the interference from missing modalities. Extensive experiments on BraTS 2018 and BraTS 2020 demonstrate that HGMamba achieves comparable Dice performance to the strongest competitor while substantially improving HD95. Furthermore, it achieves these improvements with an order of magnitude fewer parameters and over 2 faster inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Learning, Missing Modality Learning, State Space Models, Graph Neural Networks, Medical Image Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X932OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces HGMamba, which represents MRI modalities and tumor regions as two types of nodes and uses Mamba to exchange information among a small number of nodes. The motivation is intuitive and reasonable: different tumor regions benefit from different imaging modalities. However, I have some concerns about the evaluation benchmarks, comparisons with other methods.

Critical strengths
  1. Clear presentation and informative figures.
  • The paper is well organized, and the figures clearly illustrate the main design choices and processing steps, making the method easy to follow.
  1. An insightful fusion strategy.
  • The method extracts features from four MRI modalities and compresses them into modality nodes. It also introduces three learnable region nodes for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET), allowing each region to draw selectively on information from different modalities. This design fits the motivation for multimodal brain tumor segmentation: using the complementary information provided by different MRI sequences. The broader idea of region-aware modality fusion could inform future work, even without adopting the specific modules proposed here.
Critical weaknesses
  1. Older benchmarks with limited data and task coverage.
  • Relying on BraTS 2018 and BraTS 2020 as the main benchmarks feels somewhat dated for a study in 2026. Both datasets focus on preoperative adult glioma segmentation and overlap substantially, with most of BraTS 2018 included in BraTS 2020. BraTS 2021 expanded the training set for this task to 1,251 cases, while subsequent BraTS challenges have covered more patient populations, tumor types, and geographic regions. Even if the scope remains limited to preoperative adult glioma segmentation, additional experiments on the corresponding BraTS 2021 training set would provide a more representative assessment on a larger, more recent benchmark.
  1. Several implementation and evaluation details need clarification.
  • The HD95 calculation, ET post-processing, and efficiency evaluation settings could affect both the interpretation of the results and the fairness of the comparisons. Specific questions are listed below.
Questions for authors
  1. HD95 calculation and paired statistical tests.
  • According to the paper and the supplied code, cases in which only the prediction or the ground truth is empty appear to be excluded from the mean HD95 without a penalty. This excludes both false positives when the ground truth is empty and complete misses when the prediction is empty. HD95 is generally undefined or treated as infinite in these situations. Please clarify this convention, provide representative examples, and report results with an explicit penalty for these cases, specifying the penalty rule.
  • Also, if different methods exclude different cases, how were the pairs constructed for the paired Wilcoxon test, and how were missing values handled?
  1. ET post-processing and its effect on the results.
  • The supplied code removes the entire predicted ET region when its total voxel count is below 500, apparently to suppress false positives. However, this step is not described in the main text. Please clarify whether the same post-processing was applied to all methods.
  • Ideally, report ET metrics both before and after this step for every method.
  1. Settings used for the efficiency comparison.
  • How was inference efficiency measured? Please specify the numerical precision, such as FP32 or BF16, and the crop sizes and sliding-window settings. Was each method evaluated using the crop size reported in its original paper?
  • For models with multiple CNN encoders processing 3D volumes, crop size and encoder width can substantially affect inference speed. Crop sizes vary across methods, for example, 80380^3, 96396^3, and 1283128^3. The base channel width also differs: mmFormer uses 8, whereas HGMamba uses 16. Parameter counts and FLOPs alone do not fully explain the resulting runtime differences. Please clarify these settings and report throughput at both the case level and the crop level.
  • The supplied code appears to run inference entirely in FP32. Some methods, such as UniME, report using efficient kernels such as FlashAttention, whose commonly used GPU implementations require FP16 or BF16. Could the authors provide two separate throughput comparisons, one in BF16 and one in FP32, evaluating all methods in both settings under a consistent protocol?
Reviewer_CLksOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This submission works on the missing modality brain tumor segmentation task, that adapts the tumor regions segmentation to different missing modalities configurations. The proposed approach is based on the graph Mamba design to enable graph (modalities and regions) interactions. It also introduce the data-driven affinity prior relating modalities to the tumor regions to quantify the contributions of these modalities to segmentation of different regions. It is evaluated that this affinity prior contributes much to the performance gain. Overall, the proposed approach achieves better overall segmentation performance.

Critical strengths
  1. The task to segment tumor regions using multi-contrast MR images with missing modalities is a fundamental task to handle realistic missing modalities cases in medical application.

  2. The proposed data-driven affinity prior is intuitive, while it improves the model's performance on the segmentation task.

  3. The proposed approach demonstrated competitive results on the 15 different configurations of missing modalities on BraTS 2018 dataset.

Critical weaknesses

I have several concerns on the novelty and experimental evaluations.

  1. It is unclear how the graph and network adapts to different missing modalities cases. For example, it should be clear how a missing modality represents and how is it (are they) learned for adaptation to different missing cases.

  2. The presentations on the nodeMamba is not clear enough, for example the specific design of U. The nodeMamba is utilized in equations (2, 3,7). Are these questions sharing the same Mamba? The writing of the formulations in section 3.2 should be improved for better understating the computational pipeline of the proposed network.

  3. It should be clarified what is the novel method contributions of the proposed approach? Overall, the utilization of mamba and network design lack sufficient contributions.

  4. In the experiments on the BraTS 2018, the Dice score is not always best among the compared methods, however the HD95 score is best for the proposed approach. What is the reasons on this discrepancy of performance across different metrics?.

  5. It is interesting to see that the affinity prior plays a central role in the performance gain as shown in Table. 5. This is a strong prior precomputed. How about training these elements in the affinity prior (may fine-tune the pre-computed affinity prior)?

  6. The ablation studies on the training loss and the associated loss weights should be conducted in the major body of the paper.

Questions for authors

As presented above, I have several concerns/suggestions on the methodology presentation and contributions, the experimental analysis of training loss, deeper analysis on the affinity prior and experimental results analysis.

Reviewer_VHnKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents HGMamba, a framework for brain tumor segmentation with missing MRI modalities. It represents imaging modalities and the nested tumor regions—whole tumor, tumor core, and enhancing tumor—as heterogeneous nodes. A fixed affinity prior, estimated from training data using Cohen’s dd, guides modality-to-region routing. NodeMamba enables node interaction, while a reliability gate and hierarchical region interaction refine the aggregated representations. The resulting information reweights and enriches convolutional features for segmentation.

Experiments on BraTS 2018 and BraTS 2020 cover all 15 nonempty modality combinations. The authors report Dice scores comparable to the strongest baseline, lower HD95, and reduced parameter counts and inference time. Additional experiments include component ablations, alternative interaction operators, and zero-shot evaluation on cross-domain data.

Critical strengths

The paper includes ablations of the affinity prior, reliability gate, and nested region interaction, alongside parameter-matched comparisons with GAT and Transformer operators. These experiments provide useful evidence about the contributions of individual components and the framework’s applicability beyond a single interaction operator.

Critical weaknesses

W1. The positioning of the central contribution is not fully aligned with the title

The title prominently emphasizes a “Graph Mamba Framework,” whereas the experimental evidence suggests that the main contribution lies in modality–region modeling and affinity-prior-guided fusion. Removing the affinity prior reduces mean Dice by 1.53 percentage points, while replacing NodeMamba with a Transformer reduces it by only 0.37 points. Although the operator-replacement experiments support the framework’s generality, they leave Mamba’s distinctive contribution unclear.

W2. Limited Dice gains and insufficient explanation of the discrepancy between metrics

The proposed method does not surpass the strongest baseline in mean Dice on BraTS 2020, achieving 77.48% versus 77.50%, despite reporting substantially lower HD95. Although these metrics measure different aspects of segmentation and need not improve together, the source of this asymmetric improvement remains insufficiently characterized.

The qualitative examples and per-case distributions provide useful observations but do not establish whether the HD95 gains arise from fewer distant false positives, improved delineation of the main lesion, or trade-offs involving undersegmentation.

W3. The source of the reported efficiency gains is not sufficiently isolated

The reported inference speedups compare complete systems with different backbones and model capacities, making it difficult to determine whether the gains arise from the proposed fusion mechanism, the lightweight encoders, or other implementation differences. The authors attribute most parameter savings to the lightweight encoders, while the graph module operates on only a few pooled nodes.

Consequently, the end-to-end measurements do not establish a distinct efficiency advantage of NodeMamba or the proposed graph formulation.

W4. Interpretability is primarily demonstrated through visualization, without sufficient faithfulness validation

Figure 6 visualizes the affinity prior and routing weights. However, their agreement is not independent evidence of explanatory faithfulness, because the prior is explicitly injected into the routing computation. Moreover, modality representations are mixed before routing, while the final spatial features are combined using a separate set of fusion weights.

The routing weights therefore cannot be directly equated with individual input modalities’ contributions to the final prediction. The current evidence mainly establishes that the fusion mechanism is inspectable.

W5. The anatomical justification for the region-interaction order lacks targeted validation

The method adopts ET → TC → WT interaction based on the containment relationship ET ⊂ TC ⊂ WT and argues that the reverse order lacks anatomical justification. However, set containment does not preclude larger regions from providing useful context for smaller ones.

Removing region interaction demonstrates the usefulness of interaction, but does not establish that the improvement arises from the prescribed hierarchical order.

Questions for authors

see Critical Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.