This research proposes and evaluates a novel deep learning model, the Inception V3 enabled Bidirectional Long Short-Term Memory Network (IV3TM), for the automated classification of three types of brain tumors, Glioma, Meningioma, and Pituitary, from Magnetic Resonance Imaging (MRI) scans. The methodology follows a sequential hybrid architecture: pre-processing, segmentation, feature extraction, and classification. The process begins by applying an Iterative Weighted-Mean Filter (IWMF) and data augmentation to enhance image quality and overcome limited training data. Next, SqueezeNet is used for computationally efficient segmentation of the tumor region. Features are then extracted by the Inception V3 network and passed to a Deep Bidirectional Long Short-Term Memory (DBLSTM) network for final classification. Experiments using the publicly available Figshare and Brain MRI datasets showed encouraging and superior results compared to existing models. The core finding is that separating the efficient segmentation task from the robust classification task significantly boosts accuracy while reducing computational complexity and model overfitting. Specifically, the hybrid approach combines Inception V3's strength in multi-scale spatial feature extraction with the DBLSTM's ability to model contextual relationships by processing image-derived features in both forward and backward directions, thereby improving the network’s understanding of spatial dependencies between regions in the image. The main points of the discussion stress that this modular strategy successfully addresses the limitations of poor performance and high complexity found in current deep learning models. This approach ultimately provides a reliable, scalable solution with enhanced generalizability and superior accuracy for precise brain tumor diagnosis.