Brain Tumor Classification Using Modified ResNet50V2 Deep Learning Model

Abstract

Accurate diagnosis and categorization of brain tumours are very necessary for establishing optimal treatment choices and forecasting the patient's probable prognosis. The histopathological analysis of biopsy specimens is still the gold standard for identifying and categorizing brain tumours in today's medical environment. The strategy that is now being used is one that is invasive, timeconsuming, and prone to human error. Because of these limitations, it is essential to use a completely automated method for multiclassification of brain tumours. The objective of this paper is to create a multi-classification of brain tumours using modified ResNet0V2 deep learning model. The accuracy of the traditional model has been improved by the addition of dropout layers, max pooling, and batch normalisation. Batch normalisation is used to normalise the activations of the preceding layer by scaling and changing their values to have zero mean and unit variance. This is accomplished by altering the scaling factor. Because of this, the impact of the internal covariate shift is reduced, training is sped up, the stability of the model is improved, and the performance of the model in generalization is elevated. The use of max pooling helps to minimise the number of parameters in the model, which in turn makes the model more computationally efficient. Max pooling also helps to improve the model's resilience to relatively minor shifts in the input. Dropout, on the other hand, helps to minimize overfitting by reducing the co-adaptation between neurons. This, in turn, forces the network to acquire characteristics that are more robust and generalizable. The proposed model was able to attain an accuracy in classification of 96.34% as a consequence of these adjustments.