Accurate surgical instrument detection and segmentation in endoscopic videos are crucial for intelligent assistance systems in minimally invasive surgery, yet many existing datasets provide only frame-level tool presence annotations, limiting spatially supervised learning. This paper introduces m2cai16-tool-seg, an enhanced extension of the M2CAI16-Tool dataset with manually annotated bounding boxes and polygon-based instance segmentation masks for seven surgical instrument classes. We further propose a lightweight YOLOv11-based framework, termed YOLOv11-DWConv, which integrates depthwise separable convolution into the backbone to reduce model complexity while preserving detection and segmentation performance. Experimental results on the proposed dataset demonstrate that the proposed approach achieves competitive accuracy with improved computational efficiency, making it suitable for real-time surgical tool analysis in resource-constrained environments.