ADVANCING BRAIN TUMOR SEGMENTATION THROUGH ENHANCED COMPUTATIONAL EFFICIENCY AND CLINICAL IMAGE ANALYSIS

Abstract

Automated brain tumor segmentation from magnetic resonance imaging (MRI) scans plays an important role in supporting accurate diagnosis, disease assessment, and treatment planning. However, achieving precise tumor segmentation while maintaining computational efficiency remains a significant challenge in medical image analysis. To address this issue, this study proposes a ResUNet50 architecture that incorporates a ResNet50 encoder into the U-Net segmentation framework. The integration of residual learning and encoder–decoder structures is intended to improve feature extraction and capture tumor regions more effectively. The proposed approach was assessed using two datasets: a Kaggle-based T1-CE MRI dataset and the BraTS 2018 dataset, allowing its performance to be examined under different imaging conditions.Experimental results demonstrate that ResUNet50 achieved Dice coefficients exceeding 0.95 and Jaccard indices above 0.91 on the Kaggle dataset, outperforming the evaluated state-of-the-art models. Further experiments on the BraTS 2018 dataset focused on whole-tumor segmentation using four MRI modalities: FLAIR, T1, T1-CE, and T2. The model demonstrated strong segmentation performance for both high-grade gliomas (HGG) and low-grade gliomas (LGG), indicating its ability to accommodate variations in tumor characteristics and MRI appearance. Moreover, paired t-tests and Wilcoxon signed-rank tests yielded statistically significant results (p < 0.05), providing evidence of improvements over the comparison methods under the evaluated experimental conditions. Beyond segmentation accuracy, the proposed architecture offers computational advantages over conventional U-Net. The reported results indicate a reduction in model parameters of more than 60% and an inference speed approximately 4.8 times faster than that of U-Net. These improvements could help reduce computational demands and make the approach more suitable for time-sensitive medical image analysis. The combination of segmentation performance and reduced computational complexity also provides a promising foundation for further investigation into practical clinical applications. Nevertheless, additional validation on independent, multi-center MRI datasets will be necessary to establish the model's robustness across different scanners, imaging protocols, and patient populations. Future research will investigate ensemble learning and bio-inspired optimization techniques to improve segmentation consistency and generalization across diverse datasets. Explainable artificial intelligence methods, including Gradient-weighted Class Activation Mapping (Grad-CAM) and saliency maps, will also be explored to visualize the image regions that influence model predictions. Such visual explanations may help clinicians better understand the model's behavior and assess its predictions alongside conventional diagnostic procedures. Overall, this research presents a computationally efficient brain tumor segmentation approach with promising experimental results, while identifying interpretability, external validation, and broader clinical evaluation as important directions for future development.