EC-YOLO: a real-time object detection framework for minute bone tumors in X-ray imaging

Accurate early diagnosis of bone tumors is crucial for optimizing clinical treatment protocols and improving patient prognosis. However, traditional manual evaluation of X-ray images is constrained by blurred lesion boundaries, high missed-detection rates for minute lesions, and insufficient adaptability to multi-scale tumor presentations. Furthermore, manual diagnostic processes lack the real-time efficiency required for clinical workflows. Consequently, these limitations elevate the risks of misdiagnosis and missed detections, rendering traditional methods inadequate for rapid clinical triage and large-scale screening demands. To address these clinical challenges, this study proposes EC-YOLO, an advanced object detection framework built upon the YOLOv11n architecture. The model integrates an Efficient Vision State Space (EVSS) module and a Context-Guided Spatial Reconstruction Feature Pyramid Network (CGRFPN) to facilitate global information capture, spatial feature calibration, and multi-scale fusion. Specifically, the EVSS module extracts global correlations and enhances local details using geometric transformations and unidirectional scanning. This maintains linear computational complexity while improving the detection of minute lesions. Concurrently, CGRFPN employs spatial self-calibration and dynamic fusion mechanisms to accurately localize regions of interest, suppress background interference, and adapt to varying bounding box dimensions. Rigorous cross-center validation on the BTXRD-2024 dataset shows EC-YOLO achieves 0.864 precision, 0.792 recall, 0.812 mAP50, and a real-time speed of 160.3 FPS. Crucially, it demonstrates a 9.4% mAP50 improvement specifically for minute lesions (< 1 cm) over the baseline. Outperforming mainstream models such as YOLOv11n and YOLOv8, EC-YOLO balances localization accuracy with computational efficiency, providing a reliable tool for rapid clinical triage and automated micro-lesion screening.