Bald Uakari Algorithm: a novel nature-inspired metaheuristic for missing pixel imputation in medical images

Abstract

Abstract This paper introduces a novel nature-inspired metaheuristic algorithm based on the social behavior and movement patterns of the Bald Uakari monkey ( Cacajao calvus ) for solving missing pixel imputation problems in medical images. The proposed Bald Uakari Algorithm (BUA) addresses the limitations of conventional imputation techniques by implementing a hierarchical multi-agent system that mimics the distinctive foraging patterns, social structures, and territorial movements of these unique primates. BUA operates through a triple-phase optimization cycle: (1) an exploration phase in which dominant agents define search regions for subordinates, (2) an exploitation phase featuring localized refinement through specialized mutation operators that preserve structural image characteristics, and (3) a collaborative convergence phase in which successful imputation patterns are shared through a weighted influence mechanism. We enhance this core algorithm with an Adaptive 8-Connected Neighborhood strategy that dynamically adjusts pixel connectivity patterns based on local image characteristics, ensuring structural coherence in reconstructed regions. Furthermore, we present a novel hybrid architecture that integrates BUA with Generative Adversarial Networks (GANs), combining the strengths of evolutionary optimization with deep learning’s generative capabilities. Extensive experimental evaluations across diverse medical imaging modalities—including brain MRI, abdominal CT, mammography, retinal fundus, echocardiography, and dental radiographs—demonstrate BUA’s superior performance compared to established metaheuristics such as Particle Swarm Optimization, Genetic Algorithms, and Ant Colony Optimization. Our approach consistently achieves higher PSNR, SSIM, and FSIM values while maintaining lower MSE scores, with all improvements being statistically significant ( p < 0.05). A detailed case study on retinal vessel segmentation further demonstrates that GAN-BUA augmentation significantly enhances segmentation accuracy across ten different U-Net architectures, with particularly pronounced improvements in pathological cases. These results establish BUA as an effective solution for medical image imputation challenges, particularly for preserving complex structures and fine details critical for diagnostic accuracy.