A GWO-Optimized UNET++ Deep Learning Framework for Brain Tumor Segmentation Using MRI Images
Main Article Content
Abstract
Brain tumor segmentation from magnetic resonance imaging is essential for accurate diagnosis, treatment planning, and disease monitoring, but lower-grade gliomas are difficult to segment because of irregular shapes, low contrast, and class imbalance. This study proposes a GWO-optimized UNet++ framework using an ImageNet-pretrained EfficientNet-B7 encoder, ResidualSkipBlocks, and spatial and channel squeeze-and-excitation attention. Patient-disjoint training, validation, calibration and test partitions, MRI-only Otsu-derived cropping, augmentation, weighted sampling, and a composite weighted binary cross-entropy and Focal Tversky loss are employed, while Grey Wolf Optimization selects key training hyperparameters. On the LGG Segmentation Dataset, the proposed framework achieved a mean per-patient Dice of 85.58% and a mean tumor-positive-slice Dice of 81.82%, with corresponding IoU values of 75.68% and 73.01%. Dataset-pooled recall and precision were 88.12% and 89.08%. These results characterize the performance of integrated framework ,while the individual contribution of its components still need to be assessed through an ablation study.