Efficient diagnosis and treatment of tumor regions depends on accurate segmentation. The automation of this task can greatly help clinical practitioners because it removes the laborious and prone to human error issues that come with manual segmentation. However, the complexity of structures, uneven tumor boundaries, and intensity changes make tumor segmentation a difficult task. We proposed an encoder-decoder-based segmentation framework built on the VGG-16 architecture as its backbone. We incorporate a Wavelet-based channel attention mechanism (WBCAM) into the skip connections to enhance feature refinement and focus on tumor-specific regions. The guided decoder (GD) employs a weighted guided loss function to address the class imbalance, emphasizing precise segmentation of smaller tumor regions. An upsampling and convolutional blocks comprise the decoder path, improving segmentation accuracy by robustly addressing tumor appearance variations. Experimental results on a publicly accessible dataset, namely, BraTS2019, demonstrate the suggested framework's capacity to increase segmentation accuracy and validate its efficacy significantly. The proposed network addresses the challenges of brain tumor segmentation by using advanced feature refinement and guided loss optimization. The combination of WBCAM and a guided decoder significantly improves segmentation accuracy, especially in detecting smaller tumor regions.