Osedax-GAN: A novel metaheuristic approach for missing pixel imputation imagery for enhanced detection accuracy of freshwater fish diseases in aquaculture

Abstract

This paper introduces Osedax-GAN, a novel framework that combines the newly introduced Osedax metaheuristic algorithm as a loss function with Generative Adversarial Networks (GANs) to address the critical challenge of missing pixel imputation in underwater imagery for freshwater fish disease detection. The research addresses a fundamental problem in sustainable aquaculture: while laboratory-based computer vision systems achieve > 95 % accuracy on high-quality images, performance degrades to 65–75 % when applied to corrupted underwater imagery characteristic of commercial environments, precisely when automated monitoring becomes most valuable. Our investigation introduces four key innovations: (1) First application of deep-sea Osedax worm foraging behavior as a biologically-inspired metaheuristic for GAN optimization, providing superior global search capabilities specifically adapted for underwater imaging challenges; (2) Novel identity block architecture that preserves critical pathological markers during restoration, ensuring disease-specific features remain detectable even with substantial image corruption; (3) Adaptive 8-connected neighborhood strategy that dynamically adjusts to variable underwater conditions including turbidity, lighting variations, and refraction effects; (4) Pathological feature loss function specifically engineered for aquaculture disease detection that enables unprecedented preservation of diagnostic markers. Comprehensive experiments demonstrate that Osedax-GAN significantly outperforms nine state-of-the-art GAN approaches across all evaluation metrics, achieving superior restoration quality (29.86 dB PSNR, 0.918 SSIM), enhanced classification performance (92.7 % accuracy, 91.7 % F1-score), and most critically, remarkable early disease detection capability (83.9 % accuracy) that enables pathogen identification 1.6 days earlier than existing methods—expanding the critical intervention window when treatment success rates exceed 85 %. The framework demonstrates exceptional computational efficiency (34.5 ms inference time, 19.7 h training) with 10.9 % faster processing and 9.9 % lower power consumption compared to the strongest baseline, making it the first practical solution for real-time disease monitoring in commercial aquaculture facilities. Statistical validation confirms significance across all metrics (p < 0.05) with large effect sizes (Cohen's d ranging from 0.74 to 2.13), while cross-validation demonstrates robust generalization across diverse underwater conditions and disease categories. The significance of this research extends beyond immediate aquaculture applications: by enabling earlier disease detection and intervention, this work directly contributes to global food security through reduced mortality rates (potentially preventing billions in annual losses), decreased antibiotic usage supporting sustainable farming practices, and democratization of advanced monitoring technology for resource-constrained facilities worldwide. Computational efficiency achievements make automated disease monitoring accessible to developing regions, while the biologically inspired optimization framework establishes a new paradigm for underwater computer vision that addresses fundamental challenges in marine environmental monitoring, establishing the foundation for next-generation autonomous aquatic ecosystem management systems.