Orthogonal Moments-Based Feature Extraction for MRI Classification

Abstract

Orthogonal moments are a current area of research in image analysis and pattern recognition. They are numerical values obtained by projecting an image intensity function onto a polynomial basis of the 2D coordinates to describe the distribution of pixels in an image space. This work proposes using orthogonal moments in MRI images as a feature extraction tool for detecting and classifying brain tumors, including gliomas, meningiomas, and pituitary cases. The method has three stages and employs the Random Forest model (RF) as its core foundation. In the first stage, Legendre Moments and the First and Second Order Chebyshev Moments are analyzed to extract features based on the weighted average of MRI image pixel intensities. In the second stage, the feature selection vector is calculated using the orthogonal moment features obtained in the previous stage. RF determines the majority vote for each class, while the Gini coefficient evaluates its concentration, leading to dimensionality reduction. In the final stage, the feature vector is utilized in a multiclass classifier framework based on RF to diagnose the type of brain tumor. The proposed methodology achieved an average accuracy of 96.49% across all brain tumor detection. Preliminary results indicate that this family of descriptors has significant potential for feature extraction in detecting brain tumors in MRI images.