<p>Earlier and precise detection of the brain tumor is crucial in enhancing patient prognosis and directing individual treatment plans. This work proposed a hybrid deep learning system combining object and pixel-level tumor detection to analyze brain tumors completely. The proposed system has three consecutive stages; tumor detection, tumor classification, and performance evaluation. The MRI images of the brain which have been stored in the database are preprocessed (resized, contrast enhancement, and edge sharpening) in the first stage which enhances the quality of the image. The boosted images are then passed through the EfficientNet model that labels them as either tumor or normal. The second stage is then used to process tumor-positive images, which are then semantically segmented with the DeepLabV3 network. This classification is aimed at dividing three tumor subregions including enhancing tumor (ET), edema (ED), and non-enhancing tumor (NCR/NET) core. Before segmentation, ground truth masks are scaled and spatially matched with the respective input images. The proposed framework is trained and validated using the BraTS 2020 dataset, which includes 369 MRI scans for training and 125 scans for validation, ensuring a diverse and clinically representative data distribution. Evaluation is conducted through cross-validation to verify the generalization and robustness of the model. At the last phase, the performance metrics between predicted segmentation masks and expert-annotated ground truth are computed by the standard performance metrics. Experimental results demonstrate that the proposed framework achieves a classification accuracy of 99.42%, sensitivity of 98.87%, specificity of 99.15%. Furthermore, segmentation metrics yield a Dice Similarity Coefficient (DSC) of 0.9953, Intersection over Union (IoU) of 0.9856, Hausdorff Distance (HD) of 0.9973, Volumetric Overlap Error (VOE) of 0.9886, and Mean Surface Distance (MSD) of 0.9960.</p>