Plant diseases can greatly impact the food chain supply and require detection at the right time to ensure healthy crops. Based on different leaf, there are certain issues when segmenting the diseases such as small spots which easily ignored, too many disease spots, indistinguishable boundaries or low-quality spots which resulted in less precise and poor segmentation, and com-plex model with high computation. To deal with these issues, MSRU-Net is constructed based on framework of U-Net but implemented Multi-Scale ResNet50 as the encoder with reduction layer to ensure compatibility of encoder with decoder for richer extraction of global and local features to better deal with various size disease. To further improve feature representation, a Spatial-Only Coordinate Space Attention (CSA) is implemented to focus more on important features and spatial awareness with light weight attention. Additionally, the skip connections are refined with Region of Interest Enhancement Plus (ROIE+) to introduce adaptive features in the skip-connection to improve the segmentation boundary. The proposed model is compared and outperformed the original U-Net, RU-Net, MRU-Net, and RCEU-Net on two datasets: PlantDoc Leaf Disease Dataset and Tobacco Dataset. Our model achieves better segmentation results on both datasets, signifying its effectiveness while maintaining efficiency.