Diabetic retinopathy (DR) is a major microvascular complication of diabetes and a leading cause of preventable blindness worldwide, necessitating accurate and early automated diagnostic solutions. Recent deep learning–based approaches have shown promising results in DR detection; however, most existing methods focus solely on disease grading and lack lesion-level interpretability, limiting robustness and clinical reliability. To address these challenges, this study proposes a multi-task learning framework termed MTL-DRNet for simultaneous diabetic retinopathy grading and lesion segmentation from retinal fundus images. The proposed architecture employs a shared convolutional backbone to learn common representations, followed by task-specific branches for DR severity classification and pixel-wise lesion segmentation. A unified multi-task loss function jointly optimizes both objectives, enabling coordinated learning of global and lesion-level features. Experimental evaluation conducted on a publicly available diabetic retinopathy dataset demonstrates that MTL-DRNet significantly outperforms single-task, lesion-based, ensemble, and attention- driven models across standard performance metrics. The proposed model achieved an accuracy of 96.2% and an AUC of 0.98, highlighting its robustness and diagnostic effectiveness. Overall, MTL-DRNet offers an interpretable, accurate, and clinically meaningful solution for automated diabetic retinopathy screening and decision support.