Brain tumor classification using hybrid spinal-EfficientNet using MRI images

Traditional approaches to brain tumor classification frequently encounter issues such as limited efficiency, class imbalance, and high computational time, which can hinder timely clinical decision-making. To address these challenges, a new hybrid framework named Spinal-EfficientNet is introduced, aiming to enhance both classification accuracy and processing speed. The proposed pipeline starts with Magnetic Resonance Imaging (MRI) brain scans obtained from a curated dataset, followed by a preprocessing stage where noise and artifacts are reduced and image quality is improved using wavelet-domain filtering techniques. Tumor segmentation is then performed using SegNet, followed by image augmentation techniques including random erasing, rotation, and shearing to strengthen model generalization. Next, significant features are extracted, encompassing texture descriptors like Angular Second Moment (ASM), contrast, sum entropy, maximal correlation coefficient, Pyramid Histogram of Oriented Gradients (PHOG), Complete Local Binary Pattern (CLBP) and statistical measures such as mean, variance, kurtosis, skewness. In the final stage, tumor classification is performed using Spinal-EfficientNet, a hybrid architecture that combines EfficientNet with SpinalNet via customized layer modifications, allowing more reliable and accurate identification of brain tumors. Experimental evaluation demonstrates that the proposed model achieves strong performance, with specificity of 92.5%, sensitivity of 92.9%, accuracy of 92.5%, and an F1-score of 91.6% under k-fold cross-validation on the BRATS 2018 dataset. The Spinal-EfficientNet framework demonstrates notable gains in accuracy when compared with existing methods. In performance comparisons, it achieves improvements of 6.05% over Convolutional Neural Network and Support Vector Machine (CNN-SVM), 4.76% over Visual Geometry Group Stacked Classifier Network (VGG-SCNet), 3.46% over Ultra-Light Brain Tumor Detection (UL-BTD), 2.92% over Adaptive Fuzzy Deep Neural Network (AFDNN), 2.70% over ResNet50 with the Enhanced Watershed Segmentation (ResNet50-EWS), 2.59% over EfficientNet-B0, and 1.41% over SpinalNet. These consistent enhancements across a range of architectures indicate its reliability and effectiveness as a strong approach for handling complex classification problems.