Enhancing Deep Learning Models for Brain Tumor Detection: The Impact of Activation Function Selection

In this study, ten different activation functions are tested for their effectiveness in brain tumor detection and classification using an available dataset in Kaggle form 5712 MRI-Images, which divide into glioma, pituitary, meningioma, and no tumor categories using a specially constructed convolutional neural network (CNN). The availability of nonlinearity in CNNs, a required component for learning intricate data patterns in medical images, is mostly dependent on activation functions. According to our findings, LeakyReLU and Tan Activation outperform the others, with the best accuracy 99.2% and significantly low loss values of 0.029 and 0.043, respectively, demonstrating their importance in the processing of brain tumor images. The findings of this comparative study reveal that the activation functions operate differently, with Mish being the least efficient. The findings emphasize the relevance of activation function selection in the construction of medical imaging models, implying that making the proper decision can significantly enhance brain tumor classification accuracy. Our findings improve CNN designs for medical image processing and provide a framework for future studies in activation function selection. The recommended strategy, which employs optimal activation functions, has the potential to improve patient care and diagnostic accuracy in cases of brain malignancies.