Epilepsy is a complex neurological condition characterized by recurrent seizures that manifest as both focal and generalized events, presenting unique challenges for diagnosis and treatment. This research investigated the impact of three distinct data selection strategies, the discrete selector technique, the continuous partial selector technique, and the complete selector technique, on electroencephalogram (EEG)-based epilepsy detection using convolutional neural network (CNN)-based approaches. Guided by the EEG continuity principle (ECP) and stochastic data reduction (SDR), we pre-processed EEG signals by normalizing, shuffling, and partitioning them into training and testing sets. The primary objective was to analyze and assess the behavior of these three strategies to recommend the best approach for brain data analysis in clinical applications. Moreover, optimizing model parameters and batch sizes led to a marked improvement in classification performance, enhancing the accuracy of epileptic seizure detection and ultimately advancing patient care outcomes. The discrete selector technique achieved an accuracy of 82.36%, precision of 83.92%, recall of 81.03%, specificity of 83.75%, F1 score of 82.46%, and an area under the receiver operating characteristic curve (ROC-AUC) of 82.39%. In contrast, the continuous partial selector technique recorded a preliminary accuracy of 66.14% in its first stage. The complete selector technique outperformed the other methods by achieving 96.62% accuracy, 94.57% precision, 96.75% recall, 94.50% specificity, 95.65% F1 score, and 95.63% ROC-AUC.

Brain tumors, though rare, are significant health risks, often reaching critical stages before diagnosis. Gliomas, classified as high-grade (HGG) and low-grade (LGG), require early detection to reduce mortality. While two-dimensional imaging has improved diagnostic techniques, three-dimensional imaging provides a more comprehensive view. This research introduces the Hierarchical Narrowing Multi-Deep Convolutional Neural Network (HNMD-CNN), a novel method for classifying brain tumors using 3D MRI images. The HNMD-CNN employs a hierarchical narrowing filtering strategy inspired by radiologists’ models. Initially, large filters identify the tumor area and extract general features, followed by smaller filters to focus on specific tumor characteristics. This approach optimizes feature extraction and representation, improving diagnostic accuracy. We conducted extensive experiments using 3D MRI images from the BraTS2018 and BraTS2019 datasets, demonstrating the HNMD-CNN’s ability to enhance convergence speed and classification accuracy without auxiliary algorithms. Our method achieved a remarkable classification accuracy of 99.93%, representing a significant advancement in 3D imaging for glioma classification. This work provides a powerful tool for early detection and accurate diagnosis of gliomas.