Brain Tumor Classification using Modified ResNet50V2 deep learning model

Abstract

Efficient diagnosis and classification of the brain tumours are essential to guide optimal treatment decisions and address the prognosis of affected patients.The primary method for identifying brain tumors is histological examination of biopsy samples, which is laborious, invasive, and prone to significant inter-observer variability. This article proposes an automated method for classifying brain tumours using the ResNet50V2 deep learning framework. By include dropout layers, max pooling, along with batch normalization layers in the design, system accuracy as well as effectiveness are further enhanced. Enhances model resilience, expedites training, and decreases internal covariate shift using batch normalization as well as activation layer. Max-pooling assists the framework grow more invariant to input translation by reducing the amount of parameters in the framework. Dropout layers, on the hand help to reduce overfitting by preventing co-adaptation of neurones which facilitates learning of more generalized features.The proposed model achieved a classification accuracy of 96.33%, which is higher than other state-of-the-art methods. Dropout layers, applied batch normalization and enabled data augmentation helped improve the generalization power of the model. Finally, comparing with baseline models from the dataset showed a substantial increase in precision, recall, and F1 score which confirmed that the model has high accuracy to detect brain tumor. These results demonstrate the effectiveness and efficiency of the proposed method in medical image classification.The precise diagnosis and categorization of brain tumors is essential for proper therapeutic planning and enhancing patients prognoses. Herein, we propose a new deep learningframework based on a modified ResNet50V2 model, which classifies brain tumor images into four groups: glioma, meningioma, pituitary tumor, and no tumor. The dropout layers help us reduce overfitting,batch normalization helps in stabilizing the training process and max-pooling helps in improving the generalization of the model. This architecture reached a classification accuracy of 96.33% on images, outcompeting other state-of-the-art methods: InceptionV3(95.95%), MobileNetV2 (93.59%), and VGG19 (90.99%). Additional evaluation showed high precision (96.41%), recall (96.11%), and an F1-score (96.10%), and a high value of Cohen’s kappa at 0.9478, showing its reliable performance acrossall tumor types. The validation and confusion matrix results showed reliable accuracy in glioma (F1-scoreof 0.98), meningioma (F1-score of 0.93), pituitary tumor (F1-score of 0.96), and no tumor (F1-score of 0.97). This demonstrates the ability of the suggested model to accurately automate braintumor diagnosis with potential clinical applicability, showing a major leap in medical image classification.