Deep learning based on unrolled algorithm has served as an effective method for accelerated magnetic resonance imaging (MRI). However, many methods ignore the direct use of edge information to assist MRI reconstruction. In this work, we present the edge-weighted pFISTA-Net that directly applies the detected edge map to the soft-thresholding part of pFISTA-Net. The soft-thresholding value of different regions will be adjusted according to the edge map. Experimental results of a public brain dataset show that the proposed yields reconstructions with lower error and better artifact suppression compared with the state-of-the-art deep learning-based methods. The edge-weighted pFISTA-Net also shows robustness for different undersampling masks and edge detection operators. In addition, we extend the edge weighted structure to joint reconstruction and segmentation network and obtain improved reconstruction performance and more accurate segmentation results.