Accurate segmentation of skin lesions from raw clinical images remains a challenging task due to noise, irregular lesion boundaries, and the lack of high-quality annotated data. In this study, we propose a hybrid segmentation pipeline that eliminates the need for manual annotation by combining U2-Net, the Segment Anything Model (SAM), and UNet. Specifically, U2-Net is employed to generate coarse lesion masks, which are then converted into bounding box prompts to guide SAM for producing refined, high-quality masks. These pseudo-labels are used to supervise the training of a UNet model without any human-annotated labels. We evaluate our framework on a curated subset of 200 images from the ISIC 2018 dataset. The trained UNet achieves a validation IoU of 0.76, a Dice score of 0.857, and a pixel-level specificity of 99.3%. This pipeline enables fully automatic and annotation-free lesion segmentation, bridging the gap between weak supervision and practical clinical deployment.