MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation

arXiv:2510.17529 · cs.CV, cs.AI, cs.LG · Submitted 2026-08-20 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation".

Jane: The paper was written by Yovin Yahathugoda, Davide Prezzi, Piyalitt Ittichaiwong, Vicky Goh, Sebastien Ourselin et al. from School of Biomedical Engineering & Imaging Sciences at King’s College London, United Kingdom and Department of Radiology at Guy’s and St Thomas’ NHS Foundation Trust, United Kingdom.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Summary: Tom: So, building on the title, MambaX-Net addresses these challenges by moving beyond single-time point models, which is what we've been talking about. They’re dealing with the fact that traditional segmentation models are often trained only on one snapshot of a patient.

Jane: Exactly, Tom. The paper explains that traditional methods fail to leverage the rich temporal information available across a patient’s imaging history when we have multiple time points for Active Surveillance. This limits their effectiveness in a real clinical setting where time is everything.

Lu: It's not just about the image but the sequence of images, and MambaX-Net is designed to handle that flow, which is crucial for understanding how tumor growth or changes in tissue affect the whole prostate gland.

Meng: The use of a semi-supervised self-training strategy is what makes this approach practical for data collection. They use pseudo-labels generated by a pre-trained nnU-Net, allowing them to train even when expert annotations are scarce.

Lalam: That self-training element means we don're not solely dependent on limited human experts; we' are scaling the potential for AI to learn from large amounts of available data, which is a huge shift in how clinical data is leveraged.

Tom: And it's not just any self-training, as they use the segmentation mask from time point t minus one to help segment at time point t. Jane, that’s a powerful way to bootstrap the learning process.

Jane: It creates a dependency where the previous scan helps guide the current one, which is incredibly smart for making consistent predictions across multiple scans of a single patient.

Lu: The concept of MambaX-Net being able to use that prior segmentation mask allows us to build on past knowledge and refine our predictions, not starting from scratch every time.

Meng: It’s a clever way to handle the scarcity of labels, making sure we can train a robust model without having thousands of experts label every patient scan.

Lalam: By allowing the model to learn from its own previous outputs, it is accelerating the path toward automated and consistent clinical decision support.

Improvements: Tom: Now, let's get into the heart of MambaX-Net—the improvements that make it so much better than existing models. The core of this architecture is the Mamba-enhanced Cross-Attention Module, or M-CAM.

Jane: Tom mentioned mamba before, but here in simple terms, it’s using this new structure to capture long-range dependencies efficiently across time points. Instead of getting bogged down trying to connect every single pixel in a giant matrix like some older models do, Mamba scales linearly.

Lu: That linear scaling is the breakthrough because when dealing with large three dee MRI volumes, traditional Transformers become computationally exhausting. Mamba manages that complexity while still being able to see the big picture across time.

Meng: The Shape Extractor Module, or SEM, takes the previous segmentation mask and encodes it into a latent anatomical representation. This is where I think we get our refinement in the boundaries of those critical zones like PZ and TZ.

Lalam: That’s a subtle but critical improvement because it means we are not just guessing where the boundary should be; we are using actual geometric features from extracting an encoded shape to guide the future predictions.

Tom: It sounds like M-CAM is effectively fusing the current image information with the previous segmentation mask in a very smart, attention-based way that captures how things have changed.

Jane: It's like taking notes from a previous meeting and applying them to this new meeting; it's using past structural knowledge to refine our current understanding.

Lu: This allows us to model the dynamic nature of the prostate rather than treating it as a static object, which is vital for accurate diagnosis in longitudinal studies.

Meng: The integration of SEM into M-CAM means we' are ensuring that while we capture temporal changes, we're also maintaining the structural integrity defined by the previous mask.

Lalam: This combination of dynamic time awareness and structural encoding is what makes MambaX-Net so much more powerful for achieving highly accurate clinical segmentation.

Conclusion: Tom: Okay, wrapping up all this research, we've seen how MambaX-Net tackles the challenges of longitudinal data using its unique components. The results show it significantly outperforms SOTA models across various training set sizes.

Jane: And it does this even when trained on limited or noisy data, which is a huge practical win for us because perfect expert labeling is rarely achievable in large clinical trials.

Lu: The fact that MambaX-Net achieves superior boundary precision in the peripheral zone, for instance, shows how effective these new architectural components are at handling complex anatomical changes over time.

Meng: From an implementation view, the efficiency metrics are also impressive; it's a very competitive model size and inference time for clinical deployment.

Lalam: The ability to use pseudo-labels means that the path to widespread adoption of automated AI in prostate cancer management is now much clearer and less dependent on resource-heavy human annotation processes.

Tom: It really feels like we are seeing a new generation of solutions here, not just incremental improvements on old U-Nets. We have to acknowledge how far these models have come.

Jane: It's exciting to see an architecture that is both robust and highly accurate, Tom, proving that combining different AI concepts can be really effective for clinical applications.

Lu: The future definitely lies in understanding how these models adapt across multiple institutions and in larger patient cohorts to confirm this generalizability.

Meng: We need to keep refining the training protocols, but it seems like we have found a very strong foundation here for automated segmentation workflows.

Lalam: For the future, this paves the way for tracking not just segmentation but potentially predicting lesion progression itself within these longitudinal studies.

Wrap-up: Tom: So, as we wrap up our discussion on MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation, I think the core message is that combining sequential processing with a specialized attention mechanism solves major problems in medical imaging.

Jane: It’s truly a powerful tool for moving towards more accurate and efficient clinical workflows. We’ve seen how it handles longitudinal data better than almost any other model we've looked at.

Lu: I remain optimistic about the possibilities, knowing that this architecture can handle complex temporal dynamics in a way that static models simply cannot.

Meng: The engineering takeaway is clear: this is a scalable, efficient solution ready to be integrated into clinical pipelines for Active Surveillance.

Lalam: It's a significant step toward automating diagnosis and ensuring better outcomes for the patients who need longitudinal monitoring.

Tom: I think that’s a perfect way to sum up the impact of MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation.

Jane: We'll be looking forward to seeing how this technology is adopted in the real world, Tom.

Yovin Yahathugoda, Davide Prezzi, Piyalitt Ittichaiwong, Vicky Goh, Sebastien Ourselin, Michela Antonelli

School of Biomedical Engineering & Imaging Sciences at King’s College London, United Kingdom · Department of Radiology at Guy’s and St Thomas’ NHS Foundation Trust, United Kingdom

cs.CV, cs.AI, cs.LG

Submitted: 2026-08-20

Updated: 2026-08-21

Comments: Added version of record from Medical Image Analysis

Journal ref: Medical Image Analysis 114 (2026) 104251

DOI: 10.1016/j.media.2026.104251

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 95/100

The gist: The following is a detailed summary of the scientific paper, MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation, extracted from the provided text: Active

Key concepts

Longitudinal MRI Segmentation
This process involves analyzing the sequence of multiple MRI scans taken over time for a single patient. Instead of treating each scan as an isolated snapshot, this approach leverages the rich temporal information and flow across all available images to accurately track changes in tissue or tumor growth.
Mamba-Enhanced Cross-Attention Module (M-CAM)
The core of MambaX-Net, M-CAM uses a specialized structure to capture long-range dependencies across time points. Unlike traditional Transformers, Mamba scales linearly, allowing it to manage the complexity of large 3D MRI volumes efficiently while still seeing the big picture.
Shape Extractor Module (SEM)
The SEM takes the segmentation mask from a previous scan and encodes it into a latent anatomical representation. This encoded shape guides future predictions, allowing the model to build on past knowledge and refine current boundaries without needing thousands of expert labels.

Terminology

Summary

The following is a detailed summary of the scientific paper, MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation, extracted from the provided text:

Active Surveillance (AS) is a critical treatment option for managing low and intermediate-risk prostate cancer (PCa), requiring accurate prostate segmentation to enable automated detection and diagnosis. However, existing deep-learning segmentation models are often trained on single-time-point datasets, rendering them unsuitable for longitudinal AS analysis, where multiple time points and a scarcity of expert labels hinder their effective fine-tuning.

To address these challenges, the authors propose MambaX-Net, a novel semi-supervised, dual-scan 3D segmentation architecture that computes the segmentation for time point t by leveraging the MRI and the corresponding segmentation mask from the previous time point. The model is designed specifically for longitudinal analysis.

MambaX-Net's novelty rests on two key architectural components:

  1. Mamba-enhanced Cross-Attention Module (M-CAM): This component integrates the Mamba block into cross attention to efficiently capture temporal evolution and longrange spatial dependencies.

  2. Shape Extractor Module (SEM): This module encodes the previous segmentation mask into a latent anatomical representation for refined zone delination.

The model employs a semi-supervised self-training strategy that leverages pseudo-labels generated from a pre-trained nnU-Net, enabling effective learning without requiring extensive expert annotations.

Methodology and Architecture:

MambaX-Net extends the 3D nnU-Net encoder-decoder architecture, utilizing dual encoders (Enct and Enct-1) to process sequential scans at time point t (I t) and time point t-1 (I t-1). The core of the integration is the M-CAM, which fuses features from I t, I t-1, and the latent representation derived from the previous mask (SEM ft-1). This fusion allows for implicit image registration in the feature space across I t, I t-1, and M t-1. The fused features are then injected into the upsampling blocks of a standard U-Net decoder (Dect).

The authors utilized a modified loss function, FT DCE (Focal Tversky Dice Cross-Entropy), which combines the Focal and Tversky losses with the DiceCE loss, designed to focus on hard negative examples and improve segmentation accuracy, particularly in irregular regions.

Experimental Setup:

The model was evaluated on two datasets: the public PI-CAI dataset (used for pre-training) and an in-house (private) AS dataset collected at Guy’s and St Thomas’ Hospital. The experimental pipeline involved first pre-training a standard 3D nnU-Net on the PI CAI training set, which provided the initial weights for MambaX-Net. Subsequently, MambaX-Net was fine-tuned using pseudo-labels generated by this pre-trained model for the in-house AS dataset.

Results:

MambaX-Net was compared against state-of-the-art baselines (nnU-Net-V2, SwinUNETR-V2, SegMamba) and the previous dual-scan model (DSM). The results showed that MambaX-Net significantly outperforms all four models, generating more accurate prostate zone segmentations even when trained on a reduced dataset with noisy labels.

Specifically, in the comparison of performance across varying training set sizes (n), MambaX-Net consistently achieved the highest Dice Similarity Coefficient (DSC) and lowest 95th percentile Hausdorff Distance (HD95) for small datasets (e.g., n=50). Furthermore, an ablation study confirmed that combining the SEM and Mamba blocks yielded the best results, while pre-registering T2W images provided no measurable performance benefit.

Conclusion:

The authors conclude that MambaX-Net offers a feasible path toward clinical translation without manual annotation or explicit image registration. The study also revealed that the performance of dual-scan models can degrade more rapidly than that of single-scan models as the volume of noisy pseudo-labels increases, indicating a vulnerability inherent in the dual-scan approach.

Improvements for AI systems

The analysis of MambaX-Net reveals several critical architectural and methodological advancements that address fundamental limitations in current longitudinal medical image segmentation. My improvements focus on generalizing these specific innovations to enhance future AI systems, ensuring robustness, temporal coherence, and data efficiency.

1. Explicit Temporal Feature Fusion (Longitudinal Integration)

  • Improvement: Future AI architectures must move beyond treating each scan as an isolated event. We must implement a Dual-Scan or Multi-Time Point framework where the current input (t) is explicitly conditioned not only on the previous image (I t-1) but also its corresponding segmentation mask (M t-1). This creates a continuous, learned mapping of temporal evolution.

  • Implementation Detail: The system must incorporate an External Semantic Guidance Block (SEM), which takes the prior mask M t-1 and processes it into a dense, latent anatomical feature vector (f t-1). This vector guides the refinement process for the current scan I t, ensuring structural consistency across time points.

2. Linear-Time Temporal Alignment (Mamba Integration)

  • Improvement: We must replace or augment traditional quadratic self-attention mechanisms with efficient, linear scaling models, such as the Mamba block, specifically within a cross-attention context. This allows the system to model long-range temporal dependencies efficiently without the computational explosion associated with standard Transformers.

  • Implementation Detail: Implement a Mamba-enhanced Cross-Attention Module (M-CAM). This module fuses features from I t and I t-1 via Mamba blocks to capture both instantaneous spatial details and the smooth, continuous flow of anatomical changes over time, effectively performing implicit feature registration in the latent space.

3. Robust Semi-Supervised Learning (Addressing Label Scarcity)

  • Improvement: The reliance on expert labels must be minimized by leveraging self-training strategies. The system should utilize a pre-trained foundational model (e a general nnU-Net architecture) to generate high-confidence pseudo-labels for large, unlabeled datasets. These pseudo-labels are then used as training supervision, allowing the the specialized MambaX component to learn from vast amounts of data that would otherwise be inaccessible due to labeling constraints.

  • Implementation Detail: The system must incorporate a Self-Training Loop where the predicted outputs from a pre-trained model are fed back into the loss function for self-supervision, allowing the fine-tuning process to proceed even on noisy or unverified data points.

4. Loss Function Optimization for Irregular Boundaries

  • Improvement: The standard Dice Loss is insufficient for complex, small, or irregular structures (like the prostate apex). We must employ a specialized loss function that focuses heavily on hard examples and boundary delineation.

  • Implementation Detail: Implement a Focal Tversky Dice Cross-Entropy (FT DCE) Loss. This combines the ability to penalize false positives/negatives (Tversky) with the ability to downweight easy examples (Focal), ensuring that computational resources are directed toward the most difficult, boundary-defining voxels.

The resulting improved AI system will possess the following specific capabilities:

  • Superior Longitudinal Segmentation: It will achieve highly accurate segmentation of dynamic or evolving anatomical structures (e.g., tracking tumor growth, monitoring tissue change) by leveraging the temporal continuity between sequential scans, far surpassing models that treat each scan in isolation.

  • Robust Performance Under Data Constraints: The system will maintain high performance and precision even when trained on small, limited datasets or datasets where expert annotations are scarce/noisy, thanks to its self-training mechanism.

  • Precise Boundary Delineation: By integrating the SEM and FT DCE loss, it will achieve superior accuracy in segmenting challenging regions—such as the apical or peripheral zones—where traditional methods often fail due to irregular geometry.

  • Efficient Deployment: Due to the linear scaling of Mamba, the system offers a high-performance alternative to quadratic attention models, enabling faster inference and more efficient deployment in clinical settings while maintaining state-of-the-art accuracy.

Abstract

Active Surveillance (AS) is a treatment option for managing low and intermediate-risk prostate cancer (PCa), aiming to avoid overtreatment while monitoring disease progression through serial MRI and clinical follow-up. Accurate prostate segmentation is an important preliminary step for automating this process, enabling automated detection and diagnosis of PCa. However, existing deep-learning segmentation models are often trained on single-time-point, expertly annotated datasets, making them unsuitable for longitudinal AS analysis, where multiple time points and a scarcity of expert labels hinder effective fine-tuning. To address these challenges, we propose MambaX-Net, a novel semi-supervised, dual-scan 3D segmentation architecture that computes the segmentation for time point t by leveraging the MRI and the corresponding segmentation mask from the previous time point. We introduce two new components: (i) a Mamba-enhanced Cross-Attention Module, which integrates the Mamba block into cross-attention to efficiently capture temporal evolution and long-range spatial dependencies, and (ii) a Shape Extractor Module that encodes the previous segmentation mask into a latent anatomical representation for refined zone delineation. Moreover, we use a semi-supervised self-training strategy that leverages pseudo-labels generated from a pre-trained nnU-Net, enabling effective learning without expert annotations. MambaX-Net was evaluated on a longitudinal AS dataset, and results showed that it significantly outperforms state-of-the-art U-Net and Transformer-based models, achieving superior prostate zone segmentation even when trained on limited and noisy data.

Sources

Related papers