Brain tumors stand as one of the most important public health matters that the world faces today, resulting in vast morbidity and mortality. Diagnosing a brain tumor timely and accurately can be crucial for formulating treatment plans. Conventional approaches to diagnosis rely on manual radiological inspection of medical images which can be relatively slow, subjective, and prone to errors. This research proposes our approach REDX-STO which is a novel ensemble-based deep learning framework integrating four state-of-the-art Convolutional Neural Networks (CNNs)—ResNet18, DenseNet121, EfficientNetB0, and XceptionNet enhanced with nature-inspired Siberian Tiger Optimization (STO) metaheuristic algorithm. Each model contributes to specific architectural advantages while STO leverages the hunting strategy of Siberian tigers to dynamically refine ensemble weights, enhancing decision robustness. YOLOv11 is utilized to segment tumor regions in MRI scans optimized with a novel hybrid algorithm (h-CHIOLOA) which uses Coronavirus Herd Immunity Optimizer (CHIO) for broad search exploration and Lyrebird Optimization Algorithm (LOA) for precise hyperparameter refinement. The segmentation of tumor regions improved classification accuracy by eliminating irrelevant information, enhancing the overall detection system. The experimental analysis demonstrates that the proposed REDX-STO framework achieved the accuracy of 99.69% surpassing most of the pre-existing models.