Objectives: Primary bone tumors such as osteosarcoma and chondrosarcoma are rare but aggressive malignancies that require early and accurate diagnosis. Although X-ray radiography is a widely accessible imaging modality, detecting small or multi lesions remains challenging. Existing deep learning models are often trained on small, single-center datasets and lack generalizability, limiting their clinical effectiveness. Methods: We propose the YOLOv11-MTB, a novel enhancement to YOLOv11 integrating multi-scale Transformer-based attention, boundary-aware feature fusion, and receptive field augmentation to improve detection of small and multi-focal lesions. The model is trained and evaluated on two multi-center datasets, including the BTXRD dataset containing annotated radiographs with lesion types and bounding boxes. Results: YOLOv11-MTB achieves state-of-the-art performance on bone tumor detection tasks. It attains a mean average precision (mAP) of 79.6% on the BTXRD dataset, outperforming existing methods. In clinically relevant categories, the model achieves small-lesion mAP of 55.8% and multi-lesion mAP of 63.2%. Conclusions: The proposed YOLOv11-MTB framework demonstrates promising generalization and accuracy for primary bone tumor detection in radiographic images. Its performance in detecting small and multiple lesions suggests potential for clinical application.