Malignant tumor identification with custom activated deep CNN architecture: TumorNet

Brain Cancer is the most dangerous type of cancer, which affects hundreds of thousands of people worldwide every year. Early identification of brain tumor is essential for timely treatment, which is crucial for the patient’s survival. Researchers have found it challenging to develop Computer-Aided Diagnosis (CAD) for Brain Tumor Identification. Many studies have discussed the feasibility of Deep Convolutional Neural Network algorithms like ResNet, Inception, VGG in Brain Tumor Classification and Detection. These Deep Learning algorithms seem to have struggled from vanishing gradient, model bias with overfitting due to smaller medical datasets available. The proposed model TumorNet adopts a custom activated Deep CNN Architecture for classifying the more dangerous malignant tumors from benevolent tumors. The proposed model is trained from scratch and able to train with a smaller dataset consisting of MRI scans. It is evaluated, and it gives an impressive accuracy of 98.24%, outperforming other state-of-the-art CNN models, ResNet-101 and EfficientNet-B0. In comparison, the F1 scores of the proposed model are estimated at 98.05%, significantly higher than the other two models, which are 82.46 and 88.33%, respectively.