<p>Accurate brain tumor detection is vital in medical diagnosis, but often demands expert clinical judgment. To support radiotherapists and improve early identification, automated methods have gained importance. However, existing approaches struggle with accuracy and real-time efficiency. This study proposes a hybrid ResneXt-MobileNet architecture, combining MobileNet’s lightweight design with ResNeXt’s powerful feature representation, achieving high diagnostic precision with reduced computational complexity. The process begins with acquiring brain images from the dataset, which are then subjected to a preprocessing pipeline. This includes contrast enhancement using histogram equalization to improve visual quality and feature visibility. Subsequently, tumor segmentation is performed using the Dense-Res-Inception Network (DRINet), a composite Deep Learning (DL) framework that synergizes the dense connectivity of DenseNet with the multi-scale feature extraction capabilities of Inception-ResNet, enabling precise delineation of tumor regions. Following the initial segmentation, the images undergo augmentation to enhance variability and improve model generalization. Feature extraction is then performed, capturing a diverse set of statistical descriptors such as mean intensity, contrast distribution, skewness, and kurtosis. In addition, texture and frequency-based characteristics are derived using Discrete Wavelet Transform (DWT) combined with Local Directional Number Patterns (LDNP), enriching the feature space. These extracted features are subsequently fed into the proposed ResneXt-MobileNet architecture, which serves as the final classification model for accurate brain tumor detection. The proposed ResneXt-MobileNet model is benchmarked against leading approaches such as Parallel Deep Convolutional Neural Network (PDCNN), You Only Look Once version 7 (YOLOv7), 2D Convolutional Neural Network (2D CNN), Whale Harris Hawks Optimization Deep CNN (WHHO-Deep CNN), EfficientNet-B0, and the Brain Tumor Classification Model based on CNN (BCM-CNN), showing superior classification performance with sensitivity of 0.968, accuracy of 0.918, specificity of 0.948, precision of 0.939, and F1-score of 0.953 on the BraTS dataset using k-fold evaluation. For segmentation, the DRINet model achieved high scores Dice of 0.968, an IoU of 0.958, a sensitivity of 0.972, a specificity of 0.967, and a low Hausdorff Distance of 1.366, demonstrating its effectiveness across multiple images. The proposed method demonstrates both technical strength and clinical relevance, providing reliable support for early brain tumor detection. It achieved a fast computational time of 6.676&#xa0;s and outperformed traditional models, PDCNN, YOLOv7, 2D CNN, WHHO-Deep CNN, EfficientNet-B0, and BCM-CNN with classification accuracy improvements ranging from 1.31% to 6.51%.</p>