Ensemble Learning for Glioma Grading: Multi-Modal Feature Fusion Using ResNet50 and EfficientNet Variants on MRI Scans

This study compares ensemble learning methods for grading gliomas by combining features from ResNet50 and various versions of EfficientNet on the MRI image dataset. We aim to find the best combination of these network structures to improve the precision and speed of glioma classification. We introduce a process that merges ResNet50 with EfficientNet variations (from B0 to B4) for extracting features and then apply ensemble learning with diverse classifiers. The approach includes a meticulous pre-processing phase where NIfTI files are transformed into NumPy arrays and key MRI slices are extracted for analysis purposes. Then we utilise ResNet50 and different versions of EfficientNet to capture features from the data, which are later merged using a variety of learning methods. The research assesses how well different combinations of classifiers perform, such as Support Vector Machines, Random Forests, XGBoost, and Logistic Regression. Our trials show that combining models performs better in grading gliomas compared to using individual models alone in terms of how correct it is and how well it can detect specific details of the tumor severity levels. The outcomes reveal that when we mix ResNet50 with versions of EfficientNet models, we have a combined positive impact by leveraging the advantages of each architecture. We notice that particular pairs of EfficientNet models, with ResNet50, deliver results overall hinting at a blend between model intricacy and capturing features effectively. This study adds to the area of medical image analysis by providing information on the use of deep learning ensembles for categorizing brain tumors effectively, which could enhance diagnostic precision in real-world clinical scenarios and encourage the investigation of diverse deep learning methods in medical imaging.