An Efficient Transfer Learning-based Model for Classification of Brain Tumor

Abstract

Brain tumor is ranked 12th deadliest cancerous tumor among children and adults. An early detection and identification of its type can help radiologists and medical practitioners in defining an effective treatment. Therefore, this study aims to devise an efficient but accurate human braintumor classification tool that exploits magnetic resonance imaging (MRI) to segregate glioma, meningioma, pituitary, and no tumor. For this, the study exploits a dataset of 7023 MR images and performed various image pro-processing steps to get brain image by removing unwanted noisy areas and margins. Further, it incorporates data augmentation techniques to enhance limited dataset and avoid overfitting issue. Finally, a modified pre-trained deep learning-based ResNet152V2 model is trained. Two separate experiments are conducted by training proposed model with and without augmented data. It is observed that the proposed network trained on augmented data significantly outperformed the network trained on original data by successfully distinguishing four clinical states of brain tumor with an overall accuracy of 98.9%.