A Robust Multiclass Classification Framework Based on Big Transfer for Brain Tumor

Abstract

Automated spatial feature extraction is crucial in classifying brain tumors into multiple classes. Using pre-trained models in training deep neural networks for classification improves sampling efficiency and simplifies hyperparameter tuning. The proposed system employs the Big Transfer (BiT) technique to pre-train the ResNetV2-50×1 model. Subsequently, this approach performs multi-class classification in investigating brain tumors and their classes. The BiT has been employed due to its emphasis on pre-training on large datasets to develop strong general visual representations. Although the method is simple, it trains the brain tumor dataset effectively by providing a diverse set of pre-trained characteristics. The proposed framework provides an enhanced methodology for classifying brain tumors. The dataset in this suggested framework is Figshare, a publicly accessible dataset comprising 3064 photos. This study employs this dataset to classify classes: Class 0 - Meningioma, Class 1 - Glioma, and Class 3 - Pituitary. The Class 4 - no-tumor dataset comprises 4010 pictures from publicly available Br35H. This work utilized a ratio of 80:20 for training and testing the framework. The features were obtained using a fine-tuned ResNetV2-50×1 model pre-trained using BiT. The classification has been conducted using different machine-learning classifiers. XGBoost has achieved the highest performance with an accuracy of 96.42% and an AUC score of 99% in Class 0, as well as 100% in Class 1, Class 2, and Class 3.