Accurate classification of brain tumors in medical imaging is considered essential for reliable diagnosis and the formulation of effective treatment strategies. This paper presents a Multi-model Feature Fusion with Transfer Learning method MFFTL, a robust framework that enhances classification performance through a multi-stage process. Initially, input images are normalized, and tumor regions are segmented using Otsu’s thresholding technique. Deep features are then extracted using GoogLeNet and ResNet-50 convolutional neural networks. These features are fused and subsequently refined through Max-Relevance Min-Redundancy (mRMR) selection and Chaotic Enriched Northern Goshawk Optimization (CE-NGO) to ensure the retention of the most discriminative attributes. A transfer learning mechanism is employed to fine-tune a pre-trained model using the optimized features, and a Fully Connected (FC) layer is incorporated to facilitate effective training. Additionally, a Gated Recurrent Unit (GRU) is utilized to assess pathological patterns from the extracted features. The proposed MFFTL model was evaluated, and an accuracy of 99.54%, precision of 98.75%, recall of 98.89%, MAP of 98.39%, Dice coefficient of 96.76, and IoU of 97.86% were achieved. The results were found to be significantly better when compared with existing models such as SVM (90.54%), CNN (88.63%), MobileNet (82.45%), and ResNet-50 (85.65%). Furthermore, the proposed MFFTL model performed better than existing literature methods where the accuracy is 98.92%, thereby validating the superiority and robustness of the proposed MFFTL model in brain tumor classification tasks.