Wavelet-Enhanced Neural Networks for Automated Brain Tumor Detection

Abstract

Brain tumor identification and diagnosis based on magnetic resonance imaging (MRI) are of prime importance in understanding the complexities of diagnosis and treatments for these diseases. For this purpose, this research work attempts to use a deep learning based solution for multi class classifications of these tumor samples available in the brain tumor MRI data available in the Kaggle site, which includes samples of glioma, meningioma, pituitary tumors and normal samples of the human brain. Additionally, this work has used generous levels of data augmentation methods of preventing overfitting during the solving process. Contrast Limited Adaptive Histogram Equalization (CLAHE) was initially applied to improve the local contrast and highlight the subtle boundaries of tumors. The feature enhancement based on wavelet was subsequently applied to obtain details on various scales, thus enriching the features for extraction by the convolutional neural network (CNN). The experimental outcome revealed that the proposed preprocessing strategy had a substantial positive effect on the performance of the proposed approach. Prior to applying the  proposed wavelet enhancement approach, the proposed model achieved a test accuracy of 83.5\% and a Kappa value of 0.7792. However, after applying the proposed wavelet approach, there was a substantial improvement in performance, and the proposed model achieved a test accuracy of 97.4\% and a Kappa value of 0.9651, indicating that the proposed approach produced substantial increments in precision, recall, and F1 measures for all classes of tumors. Furthermore, to ensure robustness and generalization of the proposed model, a 5-fold cross-validation strategy was employed.  The results across the folds were consistent, with the best performing fold achieving an accuracy of 97.4\%, confirming the stability and reliability of the proposed approach.  Overall, the proposed method demonstrates significant improvements in precision, recall, and F1-score across all tumor classes.