One of the earliest detectable signs of diabetic retinopathy (DR)—a leading cause of vision impairment worldwide—is the presence of microaneurysm(MA). Precise segmentation and identifying very early lession like microaneurysms from retinal fundus images are critical for timely diagnosis and intervention. In the proposed research work, we have developed an enhanced deep learning framework that integrates U-Net architecture with a ResNet50 encoder for the automated segmentation of microaneurysm. The ResNet50 backbone, pretrained on ImageNet, leverages residual learning to extract fine-grained details as well as broader contextual information. The U-Net decoder reconstructs detailed segmentation masks via progressive upsampling and skip connections, preserving spatial accuracy. Model development involved training and validation on the IDRiD database [1], which contains retinal images from Indian patients, it offers finely annotated high-resolution images of retina, marking DR lesions and typical retinal structures at the pixel level. To improve generalization, we apply various preprocessing steps and data augmentation techniques during training. Quantitative evaluation demonstrates the effectiveness of our approach, achieving a Dice score of 85.69%, Intersection over Union (IoU) of 93.56%, mean accuracy of 99.55%, and mean precision of 96.01%. These results show a significant improvement over conventional U-Net models in accurately segmenting small and scattered lesions. This study demonstrates the effectiveness of integrating deep residual networks, within encoder-decoder frameworks and establishes a strong foundation for deploying automated DR screening tools in clinical environments.