Efficient Brain Tumor Classification on Resource-Constrained Devices Using Stacking Ensemble and RadImageNet Pretrained Models

Abstract

The classification of brain Tumors relies to a large extent on the knowledge and skills of experts. The introduction of a Tumor Classification System is essential to help radiologists and physicians identify tumors. It is essential to improve the accuracy of the systems in order to guarantee possible therapeutic approaches. In this study, we provide a portable system for classifying Magnetic Resonance (MR) brain images into normal and tumor categories using a Raspberry Pi 4 Model B. We employ a stacking ensemble to construct an ensemble classifier, with logistic regression acting as the meta-learner. Our approach utilizes a pre-trained Convolutional Neural Network (CNN) model on RadImageNet, specifically Inception-V3, as the feature extractor, while SVM, k-NN, and RF serve as base learners. The dataset used in this study comprises 3000 Tl-weighted MR brain images. Our method improves accuracy while addressing resource constraints. The practical implementation of the Raspberry Pi 4 Model B emphasizes real-world applicability. With a 98.5% accuracy, this approach contributes to brain tumor detection, assisting in correct diagnosis and viable therapy solutions.