Abstract: Addressing the challenges of traditional anchor worm detection methods for black bass—which are time-consuming, require specialized personnel, and face scalability issues—this study aims to develop a rapid and accurate deep learning detection technique for anchor worms. It seeks to clarify the performance differences among multiple mainstream object detection algorithms, refine and optimize the preferred model, and provide technical support for parasite identification and disease prevention in black bass aquaculture. To address the limitations of traditional detection methods, establish a rapid and accurate detection approach for Ancyrocephalus percae, and enhance the efficiency of disease prevention and control. A total of 286 microscopic images of A. percae were collected through in vivo dissection experiments, and expanded to 1,028 training images via preprocessing such as data augmentation for training deep learning networks. To select a suitable base model, the performance differences among three mainstream object detection algorithms (YOLOv11, SSD, and Faster-RCNN) were systematically compared. Aiming at the problems of blurred edges, scattered focus, and background interference in the heatmap of the selected YOLOv11 model during detection, two core innovative improvements were proposed: first, the dynamic convolution module, which replaces traditional static convolution, dynamically adapts the convolution kernel parameters of input features through a generation network, dynamically adjusts weights according to the edge texture of the parasite and details of tiny targets, and enhances the refinement of feature perception; second, the Dual Branch Block, which introduces parallel feature processing branches, with one focusing on extracting local details of high-resolution tiny targets and the other strengthening the suppression of global background interference, realizing multi-scale feature complementation through a lightweight fusion mechanism. The results showed that YOLOv11 took the lead in all key performance indicators, with a map@0.5 value of 96.6%, precision of 96.9%, recall of 91.1%, and FPS of 86 frames per second; in contrast, the map@0.5 values of SSD and Faster-RCNN were 81.19% and 79.25% respectively, precision was 93.33% and 55.79% respectively, and recall was 63.64% and 80.91% respectively, all performing inferior to YOLOv11. However, YOLOv11 had issues with its heatmap, including blurred edges, unfocused focus, and background interference. The optimized YOLOv11-DD achieved a detection speed of 72 frames per second, meeting the demand for rapid detection, and its accuracy and map@0.5 were improved to 97.1% and 98.2% respectively. YOLOv11-DD, developed through in-depth optimization of YOLOv11, effectively addresses the heatmap issues of the original model, balances detection speed and accuracy, and enhances the rapid diagnosis capability of A. percae disease. This study not only provides a new technology for the diagnosis of fish parasites but also offers a new approach for the prevention and control of fish parasitic diseases.