Related Experiment Videos
GDNet: A Robust 2.5D Multimodal MRI Brain Tumor Segmentation Framework with EMA Stabilization and Tumor-Aware
Behnam Kiani Kalejahi1,2, Sajid Khan1, Mohammad Javad Rajabi2
1Department of Computer Science, School of Engineering, Central Asian University, Tashkent 111211, Uzbekistan.
Journal of Imaging
|July 27, 2026
Summary
GDNet, a 2.5D multimodal MRI segmentation framework, achieves high accuracy in adult glioma delineation using efficient training strategies. This approach offers a practical alternative to computationally intensive 3D models for clinical applications.
Area of Science:
- Neuro-oncology
- Medical Image Analysis
- Artificial Intelligence in Medicine
Background:
- Accurate segmentation of adult diffuse gliomas from multi-parametric MRI (mpMRI) is crucial for quantitative neuro-oncology.
- Current 3D volumetric networks achieve high performance but demand significant computational resources (GPUs, long training times).
- 2D slice-wise models lack inter-slice context, hindering the segmentation of thin tumor structures.
Purpose of the Study:
- To introduce GDNet, a computationally efficient 2.5D multimodal MRI segmentation framework for adult gliomas.
- To evaluate GDNet's performance on the BraTS 2024 cohort, comparing it against existing methods.
- To identify key training strategies contributing to segmentation accuracy and computational efficiency.
Main Methods:
- GDNet utilizes a compact 2.5D U-shaped encoder-decoder architecture processing three adjacent MRI slices (12 channels total).
- Key training components include Exponential Moving Average (EMA) weight stabilization, tumor-aware slice sampling, and a compound Cross-Entropy + Soft-Dice loss.
- A systematic ablation study assessed the impact of various components, including EMA, sampling, loss functions, and input context window size.
Main Results:
- The final 3-slice 2.5D GDNet achieved high positive-only Dice scores on the BraTS 2024 test set: 0.791 (WT), 0.736 (TC), and 0.654 (ET).
- Mean foreground Dice exceeded 0.927, demonstrating strong overall segmentation performance.
- Ablation studies identified EMA + mixed sampling and the 2.5D context window as primary drivers of performance improvement.
Conclusions:
- GDNet effectively recovers substantial accuracy of heavier 3D networks with significantly reduced computational cost.
- The framework demonstrates reproducible results across seeds and outperforms a heavier GDNet-inspired architectural variant.
- GDNet presents a practical and potentially clinically deployable solution for multimodal glioma segmentation on standard hardware.