GA-UNet: Genetic Algorithm-Optimized Lightweight U-Net Architecture for Multi-Sequence Brain Tumor MRI Segmentation

Brain tumor segmentation plays a critical role in accurate diagnosis and treatment planning but remains challenging due to complex tumor boundaries and variations across MRI sequences. Although U-Net is a widely adopted architecture for medical image segmentation, its performance heavily depends on manually defined configurations, which are often suboptimal and labor-intensive. This study proposes GA-UNet, a Genetic Algorithm (GA)-assisted framework that automatically optimizes U-Net architectures for multi-sequence brain tumor MRI segmentation. GA-UNet dynamically evolves encoder, bottleneck, and decoder configurations to balance segmentation accuracy and computational efficiency. The framework is evaluated on two datasets: BraTS 2021, comprising four MRI sequences (T1, T1Gd, T2, FLAIR) for whole-tumor segmentation, and Figshare for multi-class tumor segmentation. GA-UNet achieves a Dice Similarity Coefficient (DSC) of 0.9121 on BraTS 2021, outperforming U-Net AG-CHprep (0.9095) and nnU-Net (0.8900). On Figshare, it yields DSC values of 0.9369 (Meningioma), 0.9202 (Glioma), and 0.9226 (Pituitary), with corresponding Jaccard Index (JI) scores of 0.8821, 0.8531, and 0.8582. These results confirm GA-UNet’s robustness in accurately segmenting complex tumor regions across both whole-tumor and multi-class tasks. Additional evaluation using the 95th percentile Hausdorff Distance (HD95) and statistical significance testing further validates its performance. Future work will explore 3D segmentation, attention mechanisms, and the integration of more advanced neural architecture search techniques.