🚀 EvoBrainNet: A Multi-Objective Evolutionary Neural Architecture Search with Self-Adaptive Mutation for Volumetric Brain Tumor Segmentation
🚀 Accurate, Efficient 3D Brain Tumor Segmentation in MRI using Evolutionary Architecture Search
📌 Full code and pretrained models will be released soon!
Precise segmentation of brain tumors in volumetric MRI is challenging due to significant heterogeneity in tumor shape, size, and intensity. EvoBrainNet tackles this by combining:
- ExoFeature Module: Enhanced contextual encoding for robust feature extraction.
- Dilated Residual Attention Pyramid (DRAP): Multiscale residual attention and channel recalibration.
- RefineUp Module: Decoder-side refinement with attention-guided upsampling.
A multi-objective evolutionary neural architecture search (NAS) framework with a self-adaptive mutation strategy jointly optimizes both segmentation quality (Dice similarity coefficient) and model efficiency (parameters, GFLOPs). EvoBrainNet outperforms nine state-of-the-art methods on multiple benchmarks and generalizes well across unseen datasets.
✅ State-of-the-Art Accuracy: Achieves 95.56% Dice and 1.42mm HD95 on BraTS 2021.
✅ Multi-Objective Optimization: Simultaneously maximizes accuracy and efficiency.
✅ Self-Adaptive Evolutionary NAS: Automatically explores optimal architectures with dynamic mutation rate.
✅ Generalization: Robust performance on BraTS 2020 and MSD Brain Tumor datasets.
✅ Ablation Proven: Each module’s effectiveness is statistically validated.
| Dataset | DSC (%) | HD95 (mm) |
|---|---|---|
| BraTS 2021 | 95.56 | 1.42 |
| BraTS 2020 | 93.08 | 1.97 |
| MSD Brain | 93.79 | 1.64 |
Outperforms 9 SOTA methods in both accuracy and efficiency.
- Modular 3D Supernet: Flexible, scalable architecture search space.
- Self-Adaptive Mutation: Dynamic evolution based on real-time performance.
- Clinical Efficiency: Optimized for parameter count and computational cost (GFLOPs).
- 🔜 Full code, pretrained models, and detailed usage instructions coming soon!
- 💡 Open source for the research and clinical community.