Deep learning for brain tumor segmentation-comparison of architectures & activation functions

The identification of Brain tumor is one of the challenging tasks in medical imaging that requires high precision for effective diagnosis and treatment planning. This paper presents a study on deep learning approaches for analysing, comparing segmentation and performance across multiple architectures and activation functions. We systematically assess 3D U-Net, ResNet50, VGG-19, and modified DenseNet121 models on BraTS (2018-2021), Figshare, and Br35H datasets. Our investigation extends to different activation functions in transfer learning configurations. Experimental results demonstrate that the study compares five activation functions ReLU, GELU, Swish, Tanh, and SELU to determine their impact on model efficiency. Results show that 3D U-Net achieves strong segmentation performance (Dice scores: 0.81 to 0.85), and VGG-19 excels in classification (99.9% accuracy on Figshare data). This study provides actionable insights for selecting optimal architectures and activation functions in clinical applications, while identifying some of the challenges and observations for future work in field of brain tumor analysis.