The segmentation of lesions such as retina edema, sub-retinal fluid and pigment epithelial detachment in optical coherence tomography (OCT) images is a crucial task for automated diagnosis of diabetic retinopathy. However, the multi-class lesion joint segmentation is very challenging due to the blurred boundary, complex structure, influence of noise, and the imbalanced class. In this paper, we propose a novel convolutional neural network with an encoder-decoder structure to perform joint segmentation of these three lesions. Unlike the common skip-connection employed in U-shape network for obtaining rich information from encoder feature map, we explore an encoder-decoder attention module (EDAM) via low-complexity non-local operation to capture more useful spatial dependency information between encoder feature and decoder feature. In this way, the network will take full advantage of the correlation information of the same stage feature and pay more attention to lesion areas. In order to capture large receptive fields and accurately segment small lesion, the modified lightweight residual network with dilated convolution is employed in encoding path. Besides, a hybrid loss, consisting of cross-entropy loss and multi-class Dice loss, is used to optimize our network. The proposed method was evaluated on a public database: AI-challenger 2018 for automated segmentation of retinal edema lesions, and achieved a compelling performance with less parameters compared to state-of-the-art networks.