Neuroblastoma and paediatric sarcomas frequently present with distant metastases at diagnosis, making accurate staging critical for treatment planning. Manual staging based on MRI and CT is time‑consuming and subject to interobserver variability. A multimodal artificial intelligence system was developed for automated staging of neuroblastoma (n=50) and sarcomas (n=34) in children. The system comprises three deep learning modules: a U‑Net for tumour segmentation on MRI, a YOLOv8‑based detector for metastatic lesions on CT, and a RuBERT natural language processing module for extracting staging information from Russian‑language radiology reports. The reference standard was multidisciplinary tumour board decisions. Performance was evaluated on a single‑centre retrospective cohort of 84 patients aged 6 months to 17 years. The segmentation module achieved a mean Dice coefficient of 0.79 for neuroblastoma and 0.81 for sarcomas. The metastasis detector showed a sensitivity of 0.95 and a specificity of 0.85 for lung nodules. The integrated classifier correctly assigned the stage in 78 of 84 cases (overall accuracy 92.8%). The full analysis pipeline took 8 minutes per patient on average, compared with approximately 45 minutes for manual reading. Staging errors (6 cases) were mainly due to occult bone marrow micrometastases not visible on imaging or postoperative findings. Surgeons agreed with the AI‑generated 3D tumour models in 92% of cases. The proposed multimodal AI system achieved expert‑comparable accuracy for automated staging of neuroblastoma and paediatric sarcomas while significantly reducing analysis time. Limitations include the modest sample size, single‑centre design and imperfect specificity (0.85).
