Brain Tumor Detection Based on Deep Features Concatenation and Machine Learning Classifiers With Genetic Selection
Mohamed Wageh, Khalid Amin, Abeer D. Algarni, Ahmed Hamad, Mina Ibrahim
Abstract
The development of brain tumors is often a result of cellular abnormalities, making it a leading factor contributing to mortality among both adults and children on a global scale. However, early detection of tumor can potentially prevent millions of deaths. In this regard, Magnetic Resonance Imaging (MRI) has become a pivotal tool for early brain tumor detection, It holds a vital significance role in enhancing tumor visibility that facilitates subsequent treatment planning and intervention. This research focuses on early stage brain tumor detection, proposing a Computer-Aided Detection (CAD) system that leverages MRI. Utilizing transfer learning, multiple pre-trained deep convolutional neural networks namely VGG-16, Inception V3, ResNet-101, and DenseNet- 201 are used to extract deep features from brain MRI images. Subsequently, the extracted deep features are concatenated and subjected to a genetic algorithm, acting as a technique for feature selection to determine the most important features. These features undergo evaluation using various machine learning classifiers. Two open-access brain MRI datasets, Navoneel brain tumor and Br35H Brain Tumor Detection datasets, are employed to assess model performance. Multiple experiments were conducted using the two datasets: one without feature concatenation or selection, and the other with both processes applied. The experimental results demonstrate that combining and selecting deep features leads to a substantial performance improvement, achieving an accuracy of 99.7% and 99.8% for the first and the second datasets, respectively, that surpasses the other methods.