GyroLCE-Seg: Accuracy–Efficiency-Balanced Visual State-Space Segmentation of Gyrodactylus Hard Structures

Fine-grained segmentation of Gyrodactylus opisthaptoral hard structures is challenging because of small targets, weak boundaries, local deformation, and debris interference. Using 175 microscopy images at a resolution of 2048×2048, we constructed a four-class pixel-level dataset and designed task-specific augmentation strategies. We propose GyroLCE-Seg, a framework designed to balance segmentation accuracy and computational efficiency by combining visual state-space modeling with LocalConv-based local-detail refinement, content-aware upsampling, deformable convolution, and efficient multi-scale fusion. Under image-level five-fold cross-validation and the four-class foreground macro protocol, GyroLCE-Seg achieved a MeanClassPrecision of 0.7336, a MeanClassRecall of 0.8746, a MeanIoU of 0.6582, and a MeanDice of 0.7886. At an input size of 1024×1024, it used 13.02 million parameters and 107.79 GFLOPs and achieved 34.0036 ± 0.4242 FPS on the reported GPU platform. Among the evaluated models, GyroLCE-Seg achieved the highest foreground macro recall and outperformed the lighter YOLO-series baselines on all four foreground macro metrics while offering lower resource consumption and higher throughput than the U-Net series. These results position GyroLCE-Seg as a practical intermediate option for biological laboratories equipped with routine workstation or mid-range GPU resources.