Bone sarcoma is a high grade underlying bone malignancy in teenagers and young adults, where the proper delineation of tumor boundary is imperative in the early diagnosis and treatment planning. The common drawback of existing segmentation techniques is that cannot handle heterogeneous tumor textures, low contrast and lesion margins, resulting in uneven clinical results. The research proposed the Fractal-Adaptive Graph Attention Segmentation Network (FAGAS-Net) and which incorporates multi-step denoising, feature improvement, and graph attention. The first step is to normalized the image and enhance its contrast in order to equalize the intensity and enhance the visibility of the tumor region. Multi-scale convoluting and fractal-based feature enhancement are then used which detects not only the global structural features but also fine tumor textures and channel attention identifies discriminative tumor regions. The improved features are used with adaptive graph attention to learn the spatial dependencies which results in a probability map, which is threshold to give the final binary segmentation mask. It has been experimentally evaluated that FAGAS-Net has better performance than other available ones with Dice Similarity Coefficient of 0.978, which represents correct and reliable results in delineating tumors. The solution creates a potent instrument to analyze bone sarcoma and forms a basis of future interconnection with multimodal imaging and processes.