Semantic segmentation of brain tumors plays a crucial role in assisting treatment by precisely delineating tumor boundaries in brain images. This aids clinicians in formulating surgical plans and targeted therapies, ultimately enhancing patient outcomes. Automatic brain tumor segmentation poses a significant challenge due to its complexity and resource-intensive nature, particularly in low-resource settings where access to high quality data and expertise is limited, hindering effective medical image analysis and treatment planning. In this study, the complex multi-label 3D segmentation problem was simplified into more manageable 2D single-label segmentation tasks and a U-Net architecture was adapted, with the integration of a pre-trained VGG19 model to reduce computational demands while effectively extracting features. Six models were trained in total, three using the conventional U-Net and three employing the VGG-infused U-Net (VIU-Net). The Experiment was performed using a total of 1424 magnetic resonance images. The study shows that the VIU-Net model achieved lower dice losses of 0.188, 0.068, and 0.106 for the segmentation of Necrotic Core, SNFH, and Enhancing Tumor regions respectively, showcasing the effectiveness of our method and it`s suitability for implementation in low resource settings.