Efficient Lesion Detection in Hysteroscopic Images Through Enhanced Attention Mechanisms and Directional Context Modeling.

Hysteroscopic lesion detection is clinically important for the early diagnosis of endometrial diseases. However, this task remains challenging due to the center-biased spatial distribution of lesions, anisotropic target morphology, and weak feature responses in minority categories. To address these issues, we propose YOLO-HSD, a lightweight detector built on YOLO11 with three task-oriented components. First, a center-sparse attention module (CSAM) introduces a learnable Gaussian distance prior to concentrate computation on informative central regions within the field of view. Second, a vertical-horizontal multi-kernel block (VHMB) aggregates anisotropic context with 7 × 1 vertical and 1 × 5 horizontal depthwise convolutions and softmax-weighted fusion. Third, a lightweight dynamic enhancer (LWDE) disentangles mean and variance statistics to amplify high-frequency edge responses of small lesions. On the HS-CMU dataset, the YOLO-HSD variant trained without external pretraining improves mAP@50 by 3.5 percentage points and recall by 5.5 percentage points over YOLO11n, while reducing parameter count by 7.0% to 2.40 M and model size to 4.98 MB.