Building Efficient Neural Networks For Brain Tumor Detection

Abstract

Brain tumor detection and monitoring is essential for any indicative system, as evidenced by years of research and the steady improvement of diagnostic techniques. Accordingly, treatment planning is essential to enhancing a patient's quality of life. There is an argument that deep learning could help with the difficulties of diagnosing and treating brain tumors. In this work, we introduced a hybrid deep neural network that combines state-of-the-art image enhancement methods such as contrast stretching, histogram Equalization, and logarithmic transformation with transfer learning, similar to DenseNet169 as well as ResNet149. Work provides a deep aspect of how can we improve the accuracy and efficiency of DCNN for prediction. For data selection, we create custom data which is derived from Br35H and Fig share repository, data containing benign, malignant, and normal images (596,928,364) after enhanced. Performance analyzed different scenarios different like all three enhancement algorithms data train with each neural network and evaluate performance. Performance results show the proposed work has significant improvement with Histogram equalized data with DenseNet169 which generated accuracy of 93.29%, precision of 94%, recall of 88%,  score of 93%, and loss of 20.37% which is the highest matrices over all trained neural networks in this work presented.