ADD-Net: An Adaptive Discriminative Network for Multiclass Lesion Detection in Hysteroscopic Images

Hysteroscopic lesion detection aims to automatically localize and classify lesions in hysteroscopic images, thereby facilitating intrauterine disease screening and clinical diagnosis. However, pronounced background clutter and interclass visual similarity in hysteroscopic images can impair feature discriminability. Although prevailing methods incorporate multiscale feature aggregation or local detail refinement, these operations rely on fixed processing schemes that constrain adaptive feature recalibration under varying imaging conditions, often resulting in representational instability and interclass confusion. To address this issue, we propose ADD-Net, an adaptive discriminative detection network tailored to multiclass hysteroscopic lesion detection. Specifically, a hybrid adaptive feature modulation strategy leverages an expert selection mechanism to dynamically recalibrate backbone features, improving model robustness under challenging hysteroscopic imaging conditions. A prototype confusion interaction mechanism enhances class discriminability by modeling class prototypes and interclass confusion relationships. In addition, the model uses lesion presence probabilities to improve detection stability. On the HS-CMU dataset, ADD-Net increases the three-run mean mAP@50 from 92.5% for Mamba YOLO-T to 93.3%, while reducing parameter count, computational cost, and inference latency by approximately 16.0%, 13.5%, and 33.3%, respectively. Experimental results demonstrate that ADD-Net improves detection performance while maintaining high computational efficiency, highlighting its potential for computer-aided diagnosis.