The early diagnosis of Diabetic Retinopathy (DR) would help prevent blindness, but current deep learning procedures are not very successful at detecting small lesions in retinal images. The originality of the present paper is the proposal of a hierarchical U-Net structure incorporated with two-fold attention strategies, spatial and channel, to enhance the identification of weak-grained pathological abnormalities within fundus images. The hierarchical design can achieve multi-scale feature extraction, and the attention modules can learn the relevant features of the lesion. Using large data sets such as EyePACS and Messidor-2 proves that the model suggested is superior to numerous state-of-the-art models in various assessment criteria (AUC, sensitivity, and F1-score). Visual attention maps and segmentation outputs prove the model's interpretability and effectiveness in clinical situations. These findings imply that the framework could be used in actual DR screening procedures, especially where resources are limited.