A novel Parametric Flatten-p Mish activation function based deep CNN model for brain tumor classification

The brain tumor is one of the deadliest diseases of all cancers. Influenced by the recent developments of convolutional neural networks (CNNs) in medical imaging, we have formed a CNN based model called BMRI-Net for brain tumor classification. As the activation function is one of the important modules of CNN, we have proposed a novel parametric activation function named Parametric Flatten-p Mish (PFpM) to improve the performance. PFpM can tackle the significant disadvantages of the pre-existing activation functions like neuron death and bias shift effect. The parametric approach of PFpM also offers the model some extra flexibility to learn the complex patterns more accurately from the data. To validate our proposed methodology, we have used two brain tumor datasets namely Figshare and Br35H. We have compared the performance of our model with state-of-the-art deep CNN models like DenseNet201, InceptionV3, MobileNetV2, ResNet50 and VGG19. Further, the comparative performance of PFpM has been presented with various activation functions like ReLU, Leaky ReLU, GELU, Swish and Mish. We have performed record-wise and subject-wise (patient-level) experiments for Figshare dataset whereas only record-wise experiments have been performed in case of Br35H dataset due to unavailability of subject-wise information. Further, the model has been validated using hold-out and 5-fold cross-validation techniques. On Figshare dataset, our model has achieved 99.57% overall accuracy with hold-out validation and 98.45% overall accuracy with 5-fold cross validation in case of record-wise data split. On the other hand, the model has achieved 97.91% overall accuracy with hold-out validation and 97.26% overall accuracy with 5-fold cross validation in case of subject-wise data split. Similarly, for Br35H dataset, our model has attained 99% overall accuracy with hold-out validation and 98.33% overall accuracy with 5-fold cross validation using record-wise data split. Hence, our findings can introduce a secondary procedure in the clinical diagnosis of brain tumors.