Brain tumor segmentation in Magnetic resonance imaging (MRI) is an essential job in the field of medical image analysis due to its pivotal role in its accurate and early diagnosis enabling timely therapy for patients. Hence, this paper presents a novel hybrid segmentation framework with its architecture modifying the traditional U-Net model with DenseNet121 convolutional neural network forming its backbone. This novel architecture is further leveraged with the inclusion of sophisticated self-attention mechanisms along both encoder and decoder paths to selectively capture fine-grained details for emphasizing efficient feature representations across different layers of the model. The work also analyzes integration of six additional state-of-art convolutional architectures, namely ResNeXt50, ResNet50, InceptionV3, DenseNet121, EfficientNetB0, and MobileNet into the U-Net framework for exhaustive backbone model selection and comprehensive comparisons for evaluating the efficacy of each model’s performance in the segmentation task using standard metrics IoU, F-score, Dice coefficient, and loss. The experimental results show a considerable improvement with the proposed DenseNet121 based U-Net model integrated with self-attention mechanism leading to an improvement of 7% and 5% in IoU and F-score metrics with respect to state-of-art models. The proposed framework also achieved an approximate Dice coefficient of 89%, outperforming all other modified U-Net variants. The results drawn from the experiments show that the proposed model could be promisingly used for highly accurate and efficient brain tumor segmentation, especially with the self-attention mechanism being instrumental in enhancing segmentation precision and reliability for diverse tumor morphologies.