New Mamba-based AI model segments brain tumors faster and more accurately

Researchers have unveiled a new artificial intelligence model, called DRBD-Mamba, designed to segment brain tumors in MRI scans with greater speed and robustness than existing state-of-the-art methods. The model, described in a peer-reviewed article published in Brain Informatics on 19 August 2026, was developed by Danish Ali, Ajmal Mian, Naveed Akhtar and Ghulam Mubashar Hassan.

Accurate brain tumor segmentation is significant for clinical diagnosis and treatment, but it remains challenging because of tumor heterogeneity. Mamba-based State Space Models have shown promising performance in this area, and they are computationally more efficient than many other neural architectures. However, the authors note that these models still incur considerable overhead for segmentation tasks because they compute features sequentially across multiple spatial axes.

DRBD-Mamba is a 3D segmentation model that captures multi-scale long-range dependencies with minimal computational overhead. The key innovation is the use of a space-filling curve to preserve spatial locality during the mapping of 3D features into 1D sequences. This reduces reliance on computationally expensive multi-axial feature scans, which are a major source of overhead in previous Mamba-based approaches.

To enrich the feature representation, the team introduced a gated fusion module that adaptively integrates forward and reverse contexts, along with a quantization block that improves robustness. These components work together to help the model handle the variability that characterizes real-world brain tumor imaging data.

A central contribution of the study is a new evaluation protocol. The researchers proposed five systematic folds on the BraTS2023 dataset to enable rigorous evaluation of segmentation techniques under diverse conditions, and they presented an analysis of common failure scenarios. The authors point out that the robustness of Mamba-based models across diverse BraTS data partitions had remained largely unexplored, leaving a critical gap in reliable evaluation.

On the 20% test set used by recent methods, the model achieved Dice improvements of 0.10% for whole tumor, 1.75% for tumor core, and 1.33% for enhancing tumor on BraTS 2023. Evaluations on the BraTS2020 dataset and on the proposed systematic folds further confirmed the superiority of the proposed method over existing state-of-the-art approaches, with consistent gains in tumor core and enhancing tumor segmentation.

Efficiency is where the model stands out most. According to the abstract, DRBD-Mamba achieves a 15x efficiency improvement over existing state-of-the-art methods while maintaining high segmentation accuracy, highlighting its computational advantage alongside its robustness.

The study, which is published open access, offers both a new architecture and a more rigorous way to compare segmentation models on the widely used BraTS benchmark data. By combining a dual-resolution bi-directional design with a more efficient feature-scanning strategy, the authors present DRBD-Mamba as a step toward reliable and computationally practical brain tumor segmentation.

Subject of Research: A deep learning model for efficient and robust 3D brain tumor segmentation in MRI.

Article Title: DRBD-Mamba for robust and efficient brain tumor segmentation with analytical insights

Article References: Ali, D., Mian, A., Akhtar, N., & Hassan, G. M. (2026). DRBD-Mamba for robust and efficient brain tumor segmentation with analytical insights. Brain Informatics. https://doi.org/10.1186/s40708-026-00326-w

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00326-w

Keywords: brain tumor segmentation, DRBD-Mamba, state space models, Mamba, BraTS 2023, MRI, deep learning, medical imaging, computational efficiency, Dice score, 3D segmentation, Brain Informatics