Efficient Hybrid Clustering Framework for Medical CT Image Segmentation: Comparative Evaluation with Traditional Methods

Abstract

Medical image segmentation is a critical step in computer-aided diagnosis and treatment planning. Clustering methods are widely adopted due to their unsupervised nature and ability to partition image pixels into meaningful regions without prior annotation. In this study, we conduct a comparative analysis of four traditional clustering algorithms—K-Means, Hierarchical, Gaussian Mixture Model (GMM), and Spectral Clustering—along with a proposed Hybrid method that integrates K-Means and Spectral features. Unlike most prior works that vary the number of clusters, we fix the cluster size to k = 6 across all methods to ensure consistency and highlight algorithmic differences. To improve robustness, the small dataset of five CT images obtained from a publicly available Kaggle repository was expanded using augmentation techniques, including CLAHE, flipping, gamma correction, Gaussian noise, and small-angle rotations. Performance was evaluated using standard cluster validation indices, namely Silhouette score, Calinski–Harabasz index, Davies–Bouldin index, as well as runtime, memory usage, and accuracy. The results demonstrate that while traditional methods achieve reasonable segmentation quality, the Hybrid clustering consistently outperformed others, achieving the highest accuracy (85.9%), lowest runtime (0.04 s), and smallest memory footprint (800 KB). This work highlights the potential of hybrid clustering frameworks as reliable alternatives for medical CT image segmentation.