Hierarchical attention and frequency fusion network for endometriotic lesion detection in laparoscopic images

Endometriosis is a common gynecological disorder that requires accurate detection of ectopic lesions in laparoscopic images for effective diagnosis and treatment. However, existing deep learning-based detection methods face significant challenges in identifying endometriotic lesions, particularly small lesions that occupy less than 5% of the image area, due to their diverse morphologies, blurred boundaries, and the loss of detail information in deep networks. To address these challenges, we propose HAFF-Net (Hierarchical Attention and Frequency Fusion Network), a novel deep learning framework for endometriotic lesion detection in laparoscopic images. The proposed method integrates two innovative modules into the YOLOv8 backbone: (1) the Hierarchical Lesion Attention Module (HLAM), which simultaneously captures local detail features, global contextual information, and structural characteristics through three complementary branches (Local Attention Branch, Global Attention Branch, and Structural Attention Branch); and (2) the Frequency-Aware Fusion Module (FAFM), which adaptively separates and enhances high-frequency and low-frequency components of feature maps through Fast Fourier Transform, effectively preserving critical detail information for small lesion detection while maintaining overall structural context. On the public GLENDA lesion-detection subset (373 annotated frames from 102 surgical cases), HAFF-Net is compared with plain YOLOv8 and additional detectors under patient-exclusive splits, standard COCO-style metrics, and patient-stratified five-fold cross-validation with bootstrap confidence intervals. Under the standard protocol, five-fold mAP@0.5 reaches 0.291 ± 0.037 for HAFF-Net vs. 0.238 ± 0.038 for YOLOv8n; held-out test recall remains 0.268 (95% CI: 0.184-0.356). We therefore frame the system as preliminary computer-aided decision support for expert review rather than clinically validated autonomous detection; external validation is still required.