Classification of Meningioma, Glioma, and Pituitary Tumors Using a Hybrid Algorithm Based On the Decomposition Method in MRI Brain Images

Shirzad HR, Kazemi A, Saberi F, Sepahdoost M and Zeinalnezhad M

Published on: 2022-12-29

Abstract

Background: Brain tumor image classification is an important part of clinical diagnosis and effective treatment. Radiologists may also misclassify brain tumor types when handling large amounts of data with multiple classes. Since image classification techniques can effectively improve the process of tumor diagnosis and treatment.

Objective: This study introduces a new hybrid method based on machine learning and deep learning to classify brain tumors in MRI brain images.

Methods: In this article, the T1-weighted contrast-enhanced images of 233 patients with three kinds of brain tumor: meningioma (300 slices), glioma (300 slices), and pituitary tumor (300 slices) have been used. First, the histogram equalization method was used for contrast adjustment in the image pre-processing stage. In the processing step, many features were extracted using the Gray Level Co-occurrence Matrix (GLCM) method and Convolutional Neural Network (CNN) models such as AlexNet and GoogleNet. Also, Principal Component Analysis (PCA) technique was used to reduce the dimensions of the features. Finally, 210 main features (including 100 features from AlexNet, 100 features from GoogleNet, and 10 features from the co-occurrence matrix) were obtained using the PCA technique for the classification stage. In this step, new OVO and supervised machine learning algorithms were introduced to classify the images. The use of decomposition methods, especially One-Versus-One (OVO) methods, can increase classification accuracy by separating the data into binary subsets during classification. This factor inspired the development of a new algorithm in this article in order to characterize the tumor in MRI images. Finally, we use a k-fold cross-validation (with k=10) method to separate the training and test data to evaluate the capabilities of the algorithm.

Results: The obtained results show that the accuracy of the designed algorithm using the fit ensemble classifier with the best performance in OVO and multiclass classification methods over other classifiers in brain tumor images was 99.11% and 97.44%, with a sensitivity of 99.89% and 97.52%, a specificity of 98.33% and 98.74%, a precision of 98.41% and 97.44%, a F1_score of

99.13% and 97.44%, and an AUC of 99.67% and 99.0%, respectively.

Conclusion: The achieved results of brain tumor image classification using the hybrid techniques indicates a significant potential for the medical image of the brain and enhance the classification performance, especially on multiclass datasets.