SMSC

Spinal metastases frequently occur in patients with advanced solid tumors and often necessitate comprehensive diagnostic evaluation to determine the primary tumor site. To facilitate clinical decision-making, we propose an AI-based framework utilizing non-contrast MRI to non-invasively and efficiently prioritize likely primary tumor sites. Our method consists of two key stages: (1) fully automatic segmentation of spinal metastatic lesions using a deep learning model, and (2) prediction of the primary tumor site based on features extracted from the segmented regions using a multi-branch CNN. By focusing on lesion-specific features, our method avoids irrelevant anatomical noise and improves classification accuracy. Experimental results on a retrospective dataset demonstrate high segmentation Dice scores and promising classification performance, with top-1 accuracy exceeding 55%. These results suggest the framework can reduce radiologists’ workload in identifying primary sites and improve diagnostic consistency, especially in cases with atypical imaging presentations.