<p>This paper introduces an efficient K-means Competitive Swarm Optimizer (KMCSO) based on the conventional CSO. We notice that CSO randomly selects pairwise particle individuals for the competition, while the spatial information of particles is ignored. This shortage of CSO may diminish population diversity and heighten the risk of converging to local optima. In contrast, KMCSO applies K-means clustering to partition the search space into localized regions and facilitate structured intra-cluster competition among particles with similar structures. Furthermore, optimal particles from different clusters engage in inter-cluster competition, which facilitates global information exchange. This clustering-based strategy maintains population diversity while enhancing exploration efficiency. We conduct comprehensive sensitivity analysis and comparison experiments in CEC2020 and CEC2022 to evaluate the performance of KMCSO, and the experimental results and statistical analysis confirm the scalability and efficiency of KMCSO. Moreover, we apply KMCSO to the ensemble of deep learning models for brain tumor detection, which achieves a significant improvement with 4.19% accuracy, 3.722% precision, 4.19% recall, and 4.283% F1 score against the second-best model. These findings underscore the comprehensive robustness and applicability of KMCSO in real-world scenarios.</p>