IDENTIFICATION OF BRAIN TUMOR ON MR IMAGES USING ENSEMBLE LEARNING MODELS

Abstract

Brain tumor have become a serious health concern for human beings worldwide. It’s began with an abnormal growth in the brain size. According to the recent statistics brain tumor caused 246,253 deaths globally. In 2019, the Pakistan Brain Tumor Epidemiology Study (PBTES) has reported 2,750 cases of brain tumor. Manual identification of brain tumors in MRI scans is difficult, time consuming, and subject to variable diagnosis. That's why automated computer-aided systems are important in ensuring accurate and early detection.  In the last few years, deep learning classifiers have been used for brain tumor detection, but the individual classifiers are not always consistent. To overcome this, we propose an ensemble as a hybrid approach. This approach based on five classifiers namely CNN, RF, SVM, KNN and LR. All models of machine learning are based on hybrid feature extraction to achieve better output and we use soft voting technique to combine the output of all classifiers for more reliable decisions.  In this study we use a dataset of 4600 MRI images for validation. We also include another unseen dataset. On the validation data, the top accuracy for the ensemble is 97.5%.  Experimental results on the unseen data (600 MRI images) directly show that the ensemble method is better than each individual model. The individual accuracies were: CNN 91.67%, RF 90.67%, SVM 90.83%, KNN 89.67% and LR 88.83%. The ensemble accuracy jumps to 97.17 % confirming the workability of the hybrid approach. This study shows that ensemble learning can dramatically enhance the performance of brain tumor detection, so it is a promising method that could be used in clinical decision support system.

Keywords: Pakistan Brain Tumor Epidemiology Study, Convolution Neural Network, Support vector machine, K- nearest neighbors, Gray Level Co-occurrence Matrix, K-Means Clustering.