Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MMRINet: Efficient Mamba-Based Segmentation with Dual-Path Refinement for Low-Resource MRI Analysis

Abdelrahman Elsayed, Ahmed Jaheen, Mohammad Yaqub

TL;DR

This work tackles automated brain tumor segmentation in multi-parametric MRI under resource constraints by introducing MMRINet, a lightweight U-Net–style architecture that substitutes quadratic attention with linear-time Mamba state-space blocks. It couples a Mamba bottleneck with Dual-Path Feature Refinement (DPFR) and Progressive Feature Aggregation (PFA) to maximize feature diversity and multi-scale fusion in a data-scarce SSA setting. On BraTS-Lighthouse SSA 2025, MMRINet achieves a mean Dice of $0.752$ and HD95 of $12.23$ mm with roughly $2.5$M parameters, outperforming heavier transformer-based baselines while maintaining efficiency suitable for low-resource clinics. The results demonstrate that state-space models can deliver competitive segmentation performance at a fraction of the computational cost, enabling broader deployment in SSA hospitals, with future directions including multi-task extensions and cross-dataset validation.

Abstract

Automated brain tumor segmentation in multi-parametric MRI remains challenging in resource-constrained settings where deep 3D networks are computationally prohibitive. We propose MMRINet, a lightweight architecture that replaces quadratic-complexity attention with linear-complexity Mamba state-space models for efficient volumetric context modeling. Novel Dual-Path Feature Refinement (DPFR) modules maximize feature diversity without additional data requirements, while Progressive Feature Aggregation (PFA) enables effective multi-scale fusion. In the BraTS-Lighthouse SSA 2025, our model achieves strong performance with an average Dice score of (0.752) and an average HD95 of (12.23) with only ~2.5M parameters, demonstrating efficient and accurate segmentation suitable for low-resource clinical environments. Our GitHub repository can be accessed here: github.com/BioMedIA-MBZUAI/MMRINet.

MMRINet: Efficient Mamba-Based Segmentation with Dual-Path Refinement for Low-Resource MRI Analysis

TL;DR

This work tackles automated brain tumor segmentation in multi-parametric MRI under resource constraints by introducing MMRINet, a lightweight U-Net–style architecture that substitutes quadratic attention with linear-time Mamba state-space blocks. It couples a Mamba bottleneck with Dual-Path Feature Refinement (DPFR) and Progressive Feature Aggregation (PFA) to maximize feature diversity and multi-scale fusion in a data-scarce SSA setting. On BraTS-Lighthouse SSA 2025, MMRINet achieves a mean Dice of and HD95 of mm with roughly M parameters, outperforming heavier transformer-based baselines while maintaining efficiency suitable for low-resource clinics. The results demonstrate that state-space models can deliver competitive segmentation performance at a fraction of the computational cost, enabling broader deployment in SSA hospitals, with future directions including multi-task extensions and cross-dataset validation.

Abstract

Automated brain tumor segmentation in multi-parametric MRI remains challenging in resource-constrained settings where deep 3D networks are computationally prohibitive. We propose MMRINet, a lightweight architecture that replaces quadratic-complexity attention with linear-complexity Mamba state-space models for efficient volumetric context modeling. Novel Dual-Path Feature Refinement (DPFR) modules maximize feature diversity without additional data requirements, while Progressive Feature Aggregation (PFA) enables effective multi-scale fusion. In the BraTS-Lighthouse SSA 2025, our model achieves strong performance with an average Dice score of (0.752) and an average HD95 of (12.23) with only ~2.5M parameters, demonstrating efficient and accurate segmentation suitable for low-resource clinical environments. Our GitHub repository can be accessed here: github.com/BioMedIA-MBZUAI/MMRINet.

Paper Structure

This paper contains 23 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Datasets
  4. Data Preprocessing and Augmentation
  5. Network Architecture
  6. Encoder
  7. Mamba Bottleneck
  8. Decoder with DPFR
  9. Progressive Feature Aggregation and Output
  10. Loss Function
  11. Experimental Setup
  12. Training Configuration
  13. Evaluation Metrics
  14. Baselines
  15. Results
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Overview of MMRINet Architecture. It combines a convolutional encoder with a Mamba-based bottleneck for efficient global context, and a decoder with Dual-Path Feature Refinement (DPFR) and Progressive Feature Aggregation (PFA) for multi-scale segmentation.
  • Figure 2: Key architectural components of MMRINet. (Left) The GroupMamba 3D Block models long-range spatial dependencies efficiently via group-wise state-space mixing and a lightweight modulation layer. (Right) The Dual-Path Feature Refinement (DPFR) module enhances decoder representations by combining fine-grained detail and coarse contextual features through adaptive gating and weighted fusion.
  • Figure 3: Qualitative segmentation results on BraTS-Lighthouse SSA 2025 validation cases with overlay. Colors: green (WT), blue (ET).