Purpose <p>Osteosarcoma is a type of aggressive malignant bone tumor that may cause severe pain and pathological fractures to patients. The preferred treatment for the femoral bone tumors is limb salvage surgery, which requires an accurate delineation of tumor regions from pre-operative MR and CT images. However, tumor images are characterized by high intra-class variation and low inter-class distinction, making it challenging to generate accurate segmentation.</p> Methods <p>We propose a novel plug-and-play multi-cue integration (MCI) module that leverages attention mechanisms to selectively aggregate features learned from complementary-tasks. The aim is to enforce shape-prior leveraging the existing target labels. Two types of complementary-tasks including (1) reverse distance field (RDF) regression and (2) tumor edge detection are proposed to encode the geometric properties of the femoral bone tumors. The proposed MCI module is then plugged into the decoder layers of the state-of-the-art (SOTA) nnUNet (we refer the nnUNet with the plugged-in MCI module as MCI-nnUNet) to enable shape-aware complementary-task learning for accurate tumor segmentation.</p> Results <p>We designed and conducted comprehensive experiments on an in-house dataset consisting of 75 paired CT and MR volumes of the femoral bone tumor. MCI-nnUNet achieved an average Dice similarity coefficient of 90.19±3.22%, an average Jaccard index of 82.28±5.23%, an average symmetric surface distance of 1.56±0.72 mm, and an average 95% Hausdorff Distance of 5.12±3.26 mm.</p> Conclusion <p>In summary, we developed a novel approach for automatic segmentation of the femoral bone tumors from pre-operative CT and MR images. Results obtained from the comprehensive experiments demonstrated the efficacy of the present approach. Our method achieved better results than other SOTA methods.</p>