Accurate segmentation of brain tumors from magnetic resonance (MR) images is crucial for effective diagnosis and treatment planning. However, this task is inherently challenging due to the complex nature of tumors, class imbalance issues and the necessity for precise boundary delineation. In this study, we proposed a modified U-Net model for brain tumor segmentation, which addresses these challenges by integrating loss functions during training. Our modified U-Net model combines binary cross-entropy loss with Dice loss, improving the strengths of the U-Net architecture while incorporating loss functions specifically designed to handle class imbalance. By creating a combined loss function, our approach guides the model toward accurate segmentation while mitigating the effects of class imbalance and improving boundary delineation. We conducted an empirical analysis using a Brain MRI segmentation dataset (Kaggle) of brain MR images. Our model achieves impressive segmentation quality metrics. Notably, our model yields a Dice score of 0.9759, an IoU of 0.9529, a Precision of 0.9813, a Recall of 0.9705, a specificity of 0.9895, an accuracy of 0.9827, a Jaccard index of 0.95 and an F1-Score of 0.9759. These metrics highlight the model’s efficacy in accurately delineating brain tumor boundaries from MR images. Additionally, statistical analyses including T-tests and Chi-square tests were performed to measure the significance of differences between the original and segmented images.