Transformer-based multi-scale feature fusion for real-time CT bone metastasis detection.

Abstract

Bone metastasis, a frequent complication of advanced cancers, requires early, precise detection to enable timely interventions and improve patient outcomes. Computed tomography (CT), valued for non-invasive, high-resolution imaging, is essential for identifying bone metastatic lesions. However, these small, morphologically diverse, low-contrast lesions, combined with complex tumor microenvironments and computational limitations, challenge deep learning models' accuracy and real-time applicability. We propose BM-DETR, a Transformer-based model integrating spatial-contextual enhancement module (SCEM) to enhance low-contrast lesion features through channel attention and spatial mixing, AttentionUpsample for superior multi-scale feature fusion via dual-branch upsampling, and dilated transformer attention block (DTAB) to improve contextual modeling, addressing transformer limitations in local detail capture while optimizing efficiency. Evaluated on OsteoScan dataset, BM-DETR achieves mAP50 of 0.9376, and on BMSeg dataset, mAP50 of 0.9139, surpassing state-of-the-art methods. Balancing high accuracy with computational efficiency, BM-DETR's potential for edge deployment supports early screening and intelligent diagnosis of bone metastases. This work provides a robust foundation for automated lesion detection, advancing clinical translation of diagnostic systems.