Utilization of GRN Layers in an Attention-Based U-Net for Enhanced Brain Tumor Segmentation

In recent years, the application of artificial intelligence (AI) in disease diagnosis has experienced significant advancements, enabling more accurate and rapid identification of medical conditions. One prominent area of AI application is in the analysis and processing of brain MRI images for tumor detection and classification. U-Net architectures have proven to be highly effective for this task, with numerous variations of these models being explored for segmentation purposes. Notably, attention mechanisms integrated with U-Net models have demonstrated promising results in enhancing segmentation accuracy. However, several challenges remain in this domain, including the optimization of loss functions for precise segmentation, the limited availability of medical datasets, class imbalance, and the urgent need to enhance model accuracy while maintaining computational efficiency. In this study, we address these challenges by incorporating a Global Response Normalization (GRN) layer into the attention-based U-Net model. Our proposed approach significantly improves segmentation performance, achieving a ۲۳% increase in Intersection over Union (IOU) and an ۱۱% improvement in Dice similarity score on the LGG dataset, demonstrating its superior performance compared to existing methods.

کلیدواژه ها:

Brain tumor ، Semantic segmentation ، Attention U-Net ، GRN