The rapid advancement of emerging technologies is reshaping brain tumor diagnosis and treatment planning, with a focus on precise segmentation techniques for early intervention. Manual segmentation methods face challenges due to inherent noise and intensity variations in medical imaging data. To mitigate these challenges, we propose DCRUNet++ for brain tumor segmentation. The proposed model integrates Depthwise convolutional residual module blocks to enhance information flow and gradient propagation across network layers, thereby improving feature representation. The DCRUNet++ architecture incorporates nested up-convolution operations, facilitating the propagation of semantic information from lower to higher levels of abstraction. To further optimize model training, we introduce a custom loss function that assigns higher weights to feature maps 4 and 8, prioritizing significant representations during the optimization process. Deep supervision, utilizing 8 intermediate feature maps, ensures robust training and facilitates convergence by emphasizing critical representations. Extensive experimentation with FLAIR MRI images validates the efficacy of the proposed DCRUNet++ model. Achieving a Dice coefficient of 0.9467 and a mean Intersection over Union of 0.9155, our model outperforms previous methodologies, underscoring its effectiveness in brain tumor segmentation and treatment planning.