Diabetic retinopathy (DR) is a leading cause of blindness, requiring accurate grading and lesion segmentation for effective diagnosis and treatment. Existing models often face challenges with large parameters and computational complexity, typically focusing on either grading or single-lesion segmentation. This paper investigates solving the above issues from the perspective of transfer learning, using a simple dual-branch model to accomplish lesion grading and multi-lesion segmentation tasks. Firstly, we employ fine-tuning to transfer the ResMLP-12 model to grade diabetic retinopathy lesions, achieving favorable results on our dataset while reducing training and computational costs. Secondly, we design an end-to-end U-shaped architecture for lesion segmentation tasks, where the pre-trained ResMLP-12 model serves as the encoder of our network, replacing the convolutional layers in the original U-shaped network, further reducing computational costs. To achieve more precise segmentation of lesions of different scales, we design a multi-scale pooling module (MSP) between the encoder and decoder. We train and validate our model on two datasets, APTOS 2019 and IDRiD. The experimental results prove the effectiveness of our model.