Brain tumor segmentation using a combination of graph cut and optimized U-Net with genetic algorithm: An efficient hybrid framework for enhanced medical image analysis
- 1Department of Mathematics, Faculty of Science Al-Azhar University, Assiut, Asyut, Egypt
- 2College of Engineering and Technology, American University of the Middle East, Egaila, Kuwait
Abstract
Accurate segmentation of low-grade gliomas (LGGs) in FLAIR MRI remains challenging due to low-contrast tumor boundaries, irregular morphology, and extreme class imbalance. This work presents a lightweight hybrid segmentation framework that integrates three complementary components: (1) a fast graph-based region refinement step using GrabCut applied during training only to enhance tumor–background contrast, (2) a compact U-Net architecture whose filter configuration is automatically selected using a small-scale genetic algorithm, and (3) an analytical evaluation of lightweight post-processing strategies studied separately from the final GA-optimized model. The method was evaluated on the LGG-MRI-Segmentation dataset using strict patient-level splits of 79 training, 19 validation, and 12 testing patients. The final model achieved strong slice-wise performance with a mean Dice of 0.9132 and a median Dice of 0.9855; the lower mean reflects the influence of more difficult, lower-performing slices and therefore provides a more conservative summary of overall performance across the evaluated test cases. Additional metrics include an IoU of 0.8851, sensitivity of 0.8954, precision of 0.9523, specificity of 0.9997, and an HD95 of 1.00 mm, indicating highly accurate boundary delineation. For contextual benchmarking, a 2D nnU-Net model was trained on the same dataset. nnU-Net achieved a mean validation Dice of 0.8450 and a best validation EMA Dice of 0.9134. Because these nnU-Net Dice values were obtained from its validation workflow rather than the exact slice-wise test procedure used for the proposed model, they are presented as contextual reference results rather than strictly matched test-set estimates. The proposed framework required substantially less training time under the evaluated implementations. Taken together, these findings demonstrate that a carefully coordinated combination of preprocessing, lightweight architectural optimization, and efficient post-processing can achieve strong segmentation performance on the evaluated dataset while remaining computationally efficient. These findings suggest that the proposed framework is promising for computationally constrained LGG segmentation workflows. However, broader clinical applicability requires validation on independent external cohorts before generalizability can be established.
Related
- EPRA U-Net: an efficient pyramid residual attention framework for accurate infarct segmentation in diffusion-weighted MRI.
- Brain tumor segmentation using particle swarm optimized histogram equalization and a VGG19 based U-Net
- Predicting brain tumours through transfer learning and optimized hyperparameters using the modified whale’s algorithm
- Combination of Technologies in Body Contouring: Evidence Mapping and a Decision-Making Algorithm
- Enhancing early-stage breakthrough technology identification: A CatBoost model optimized with the SABO algorithm
BibTeX
@article{mostafa2026brain,
title = {{Brain tumor segmentation using a combination of graph cut and optimized U-Net with genetic algorithm: An efficient hybrid framework for enhanced medical image analysis}},
author = {Reem M. Mostafa and Emad Mabrouk and Mohamed Saed Salim and Ahmed Ayman and Hamdy Z. Zidan},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0338773},
url = {https://doi.org/10.1371/journal.pone.0338773},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Brain tumor segmentation using a combination of graph cut and optimized U-Net with genetic algorithm: An efficient hybrid framework for enhanced medical image analysis
AU - Reem M. Mostafa
AU - Emad Mabrouk
AU - Mohamed Saed Salim
AU - Ahmed Ayman
AU - Hamdy Z. Zidan
PY - 2026
DA - 2026/10/14
JO - PLOS One
DO - 10.1371/journal.pone.0338773
UR - https://doi.org/10.1371/journal.pone.0338773
AB - Accurate segmentation of low-grade gliomas (LGGs) in FLAIR MRI remains challenging due to low-contrast tumor boundaries, irregular morphology, and extreme class imbalance. This work presents a lightweight hybrid segmentation framework that integrates three complementary components: (1) a fast graph-based region refinement step using GrabCut applied during training only to enhance tumor–background contrast, (2) a compact U-Net architecture whose filter configuration is automatically selected using a small-scale genetic algorithm, and (3) an analytical evaluation of lightweight post-processing strategies studied separately from the final GA-optimized model. The method was evaluated on the LGG-MRI-Segmentation dataset using strict patient-level splits of 79 training, 19 validation, and 12 testing patients. The final model achieved strong slice-wise performance with a mean Dice of 0.9132 and a median Dice of 0.9855; the lower mean reflects the influence of more difficult, lower-performing slices and therefore provides a more conservative summary of overall performance across the evaluated test cases. Additional metrics include an IoU of 0.8851, sensitivity of 0.8954, precision of 0.9523, specificity of 0.9997, and an HD95 of 1.00 mm, indicating highly accurate boundary delineation. For contextual benchmarking, a 2D nnU-Net model was trained on the same dataset. nnU-Net achieved a mean validation Dice of 0.8450 and a best validation EMA Dice of 0.9134. Because these nnU-Net Dice values were obtained from its validation workflow rather than the exact slice-wise test procedure used for the proposed model, they are presented as contextual reference results rather than strictly matched test-set estimates. The proposed framework required substantially less training time under the evaluated implementations. Taken together, these findings demonstrate that a carefully coordinated combination of preprocessing, lightweight architectural optimization, and efficient post-processing can achieve strong segmentation performance on the evaluated dataset while remaining computationally efficient. These findings suggest that the proposed framework is promising for computationally constrained LGG segmentation workflows. However, broader clinical applicability requires validation on independent external cohorts before generalizability can be established.
SN - 1932-6203
ER -