A hybrid deep learning framework based on VGG19 and U-Net for accurate brain tumor segmentation in MRI images.
Authors
Affiliations (3)
Affiliations (3)
- Department of Computer Engineering, Faculty of Engineering, Dicle University, 21000, Diyarbakır, Turkey.
- Department of Computer Engineering, Faculty of Engineering, Mardin Artuklu University, 47000, Mardin, Turkey.
- Department of Electricity and Energy, Silvan Vocational School, Dicle University, 21000, Diyarbakır, Turkey.
Abstract
Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) is a critical task in neuro-oncology, directly impacting diagnosis, treatment planning, and patient outcomes. This study presents a hybrid deep learning model that integrates the VGG19 convolutional neural network as encoder and the U-Net architecture as decoder for automated tumor segmentation. The model is trained and validated using the lower-grade glioma (LGG) segmentation dataset from the cancer imaging archive (TCIA), which includes annotated FLAIR MRI scans of LGG cases. By leveraging the deep feature extraction capabilities of VGG19 and the spatial precision of U-Net, the proposed model achieves superior segmentation accuracy while maintaining architectural simplicity. Quantitative evaluation demonstrates a Dice of 92.20%, intersection over union (IoU) of 85.53%, and area under the curve (AUC) of 95.90%, outperforming several state-of-the-art models in the literature. This framework offers a reliable and efficient solution for clinical MRI analysis and has strong potential for integration into computer-aided diagnostic systems. The results indicate that hybrid convolutional neural network (CNN) architectures can significantly enhance the accuracy and robustness of brain tumor segmentation tasks. The novelty of this work lies in the seamless integration of VGG19's pre-trained deep features with U-Net's skip connections, providing a lightweight yet highly accurate model that reduces computational complexity compared to multi-branch or attention-based architectures, while achieving superior performance on the LGG dataset.