ScaleMamba-YOLO: a multi-scale MambaYOLO for medical object detection

The efficacy of automated lesion detection in clinical settings is often hampered by two primary factors: the vast range of pathological scales and the presence of non-target anatomical interference. Standard Mamba-based detectors, while efficient, frequently suffer from fixed receptive fields and background signal leakage. To resolve these challenges, we introduce ScaleMamba-YOLO, an enhanced medical object detection framework that integrates selective state-space modeling with adaptive local feature refinement. Our approach features two innovative structural components: first, a Medical Multi-scale Local Feature Enhancement Block (MMLFE-Block) is positioned at the frontend to diversify the initial receptive field. By utilizing a parallel architecture with heterogeneous convolutional kernels (1×1, 3×3, and 5×5), the model achieves comprehensive hierarchical perception, enabling the concurrent identification of minute calcified spots and extensive diffuse lesions. Second, a Partial-Enhanced C2F (PEC2F) module is designed to refine feature aggregation post-global modeling. This component employs partial convolution (PConv) to selectively process salient feature channels, effectively filtering out irrelevant background noise from normal tissue structures. The robust performance of ScaleMamba-YOLO was validated across three specialized medical datasets (Br35H, BCCD, and PLoPy) and a standard scene dataset (VOC0712). The model recorded Average Precision (AP) scores of 72.7%, 65.0%, 85.7%, and 64.6%, respectively. These metrics represent consistent improvements of 1.7% to 2.3% over the MambaYOLO baseline, underscoring the system's potential for high-fidelity diagnostic assistance in real-time clinical workflows.