Brain metastasis segmentation using AI: Graylight Imaging presents new results at MICCAI 2026

For another consecutive year [1], the Graylight Imaging team had the privilege of participating in the Medical Image Computing and Computer Assisted Intervention (MICCAI 2026) conference and contributing to one of the most important international challenges in medical image analysis: BraTS-METS 2026 (Brain Tumor Segmentation Challenge – Metastases). Our team presented the poster “Brain Metastasis Segmentation Using an Ensemble of U-Mamba Models” [2], showcasing a new AI-based approach to automated brain metastasis detection and segmentation.

Beyond presenting our research, we are proud to share our results achieved in the BraTS-METS 2026 Challenge:

• 2nd place out of 35 teams in the detection task
• 4th place out of 35 teams in the segmentation task

These results reflect months of research, experimentation, and collaboration aimed at improving the accuracy and reliability of AI tools supporting neuro-oncology workflows.

We would also like to extend our sincere congratulations to all challenge organizers, contributors, and participating teams. The BraTS initiative continues to play a critical role in advancing transparent, objective, and clinically meaningful evaluation of AI algorithms in neuroimaging.

Why brain metastasis detection matters

Brain metastases are among the most common intracranial tumors and require regular monitoring throughout the patient’s treatment journey. In current clinical practice, assessment often relies on manual measurements performed according to the Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) guidelines. This process can be time-consuming, labor-intensive, and susceptible to inter-observer variability.

The challenge becomes even greater when patients present with numerous small lesions. Accurate identification and volumetric quantification of both tumors and surrounding tissue changes are crucial for:

• monitoring disease progression,
• evaluating treatment response,
• supporting therapeutic decision-making,
• ensuring consistent longitudinal assessment.

These limitations have fueled growing interest in AI systems capable of automating lesion detection and segmentation while maintaining clinical reliability.

Building on a winning foundation

Our work presented at MICCAI 2026 builds upon the Graylight Imaging solution that won the BraTS-METS 2025 challenge. Last year’s approach relied on nnU-Net, a self-configuring deep learning framework that has become one of the most successful architectures in medical image segmentation.

For BraTS-METS 2026, we explored whether emerging architectures could push performance even further.

The result is an ensemble approach based on U-Mamba, a novel architecture that combines:

• convolutional neural networks responsible for extracting local imaging features,
• state space sequence models capable of modeling long-range spatial dependencies across volumetric MRI data.

By combining these advantages, U-Mamba can capture both fine anatomical details and broader contextual information within a scan.

A special focus on small metastases

One of the most clinically important challenges in brain metastasis analysis is the detection of very small lesions.

Even tiny metastatic foci may significantly influence treatment planning, radiosurgery decisions, and patient monitoring. Missing a lesion can have direct clinical consequences.

To address this problem, our researchers incorporated the recently proposed CATMIL loss function, specifically designed to improve the detection and segmentation of small structures. CATMIL emphasizes lesion-level learning while assigning greater importance to smaller lesion components during training.

Analysis performed during model development showed that the CATMIL-enhanced U-Mamba model consistently achieved higher sensitivity when identifying small metastatic lesions.

Combining the best of both models

Interestingly, our experiments revealed that no single model dominated across all clinical scenarios.

While standard U-Mamba achieved stronger segmentation performance for larger lesions, the CATMIL variant demonstrated superior detection of smaller metastases. This observation led us to develop an ensemble strategy that combines predictions from both specialized models.

The ensemble dynamically leverages:

• standard U-Mamba for larger lesions,
• CATMIL-enhanced U-Mamba for smaller lesions,
• additional post-processing steps to reduce false-positive findings and improve resection cavity handling.

This design allowed us to balance lesion detection sensitivity with overall segmentation quality.

Real-world robustness matters

One reason BraTS-METS remains such an influential benchmark is the diversity of its dataset.

The 2026 challenge dataset contains MRI studies collected from nine institutions worldwide and includes variability in imaging equipment, acquisition protocols, image quality, and clinical settings [3].

Developing algorithms that perform consistently across such heterogeneous data is essential for creating AI solutions that can eventually support routine clinical workflows.

The challenge therefore goes beyond maximizing benchmark metrics. It focuses on building systems capable of generalizing to real-world healthcare environments.

Looking beyond segmentation

The introduction of a dedicated detection leaderboard in BraTS-METS 2026 reflects a broader trend in healthcare AI: clinicians increasingly need algorithms that not only delineate lesions accurately but also ensure that no clinically relevant findings are missed.

This shift aligns closely with our research priorities at Graylight Imaging.

Future work will focus on:

• further improving small lesion sensitivity,
• enhancing resection cavity analysis,
• increasing model robustness across institutions,
• extending AI support for quantitative neuro-oncology workflows.

Another year at MICCAI, another step forward

MICCAI remains one of the world’s most important forums for advancing medical imaging, computer-assisted intervention, and healthcare AI.

For Graylight Imaging, participating in MICCAI and BraTS has become an annual opportunity to benchmark our solutions against leading international research groups, validate our methods through objective evaluation, and contribute to the advancement of AI in neuro-oncology.

We are proud that our team’s work achieved:

✅ 2nd place in brain metastasis detection (35 participating teams)
✅ 4th place in brain metastasis segmentation (35 participating teams)

Most importantly, we are grateful to the organizers, reviewers, researchers, clinicians, and fellow participants who continue to make the BraTS community one of the most impactful collaborative initiatives in medical AI.

Congratulations to everyone involved, and we look forward to the next edition.

The poster “Brain Metastasis Segmentation Using an Ensemble of U-Mamba Models” was presented by Graylight Imaging researchers Piotr Rudzki, Maria Bancerek, and Jakub Nalepa at MICCAI 2026. The work builds upon the team’s BraTS-METS 2025 winning solution and introduces a new ensemble framework based on U-Mamba architectures for automated brain metastasis detection and segmentation.

References:

[1] Read more about our 2025’s performance.

[2] Check out our blog post on U-Mamba: https://graylight-imaging.com/blog/u-mamba-rising/

[3] https://challenges.synapse.org/Challenges/DetailsPage/Overview?id=syn74274097