ENHANCING BRAIN TUMOR CLASSIFICATION THROUGH ADVANCED IMAGE PROCESSING, HYBRID FEATURE EXTRACTION, MRMR FEATURE SELECTION, AND MULTI-CLASS CLASSIFIERS

The purpose of this research is to create a holistic methodology for detecting and classifying brain tumors using up to date image processing, feature extraction and machine learning methods. To improve the diagnostic accuracy, this research evaluated the model performance of various classifiers as well as expanding feature representation to select the best available model for Brain tumor classification. Techniques: Such process includes developing techniques of image enhancement using TV-L1 norm and MRI segmentation algorithms etc. Preparation: Using sophisticated MRI segmentation techniques, the areas of tumors were precisely identified and improved image quality by implementing the TV-L1 standard for efficient image augmentation. For robust feature extraction and feature richness, Multi Scale Local Binary Pattern (MSLBP) for texture analysis was integrated with the multi-dimensional feature representation feature through the use of Quaternion Wavelet Transform (QWT). The Histogram of Oriented Gradients (HOG) is leveraged to quickly capture edge and shape data. Selecting Features Effectively: the dimensionality is reduced and hence increased classification effectiveness by selecting and prioritizing the most useful features via the Minimum Redundancy Maximum Relevance (MRMR) technique. Different Classification Techniques: Random Forest has great accuracy and is robust, so a number of classifiers were constructed. When it comes to modeling complex data distributions, Support Vector Machines (SVMs) prove effective. The K-Nearest Neighbors (KNN) is utilized for local patterns because it is user friendly and good. Results: The dataset from China's Nanfang Hospital and General Hospital was used to evaluate the techniques and is available on Kaggle. With low error rate (0.0147), high sensitivity (98.53%), and specificity (99.51%), Random Forest (RF) had the highest accuracy of 98.53%. In the end, Naive Bayes (NB) performed the worst at 96.67% while SVM and KNN produced slightly lower accuracies at 97.22% and 97.50% respectively. More sophisticated processing, feature extraction, and selection procedures used as part of the suite of preprocessing operations contributed significantly to the increased classification accuracy. RF revealed great promise for clinical use in identifying brain tumors, achieving high sensitivity and specificity while decreasing false positives to an accuracy rate of 98.53%. Even though RF consistently outperformed the others like SVM and KNN on all measures, they did very well on their own as well.