The clustering algorithms, like is the K-means algorithm, are commonly utilized for the biomedical image regional segmentation. One of the major limitations of the clustering algorithms is a definition of the initialization phase. When the initialization distribution of the centroids is improperly set the K-means algorithm is not able to achieve a reliable approximation of the tissues, thus the convergence of such segmentation procedure is significantly limited. Furthermore, when the biomedical image data are corrupted either by the noise, or artefacts, an effectivity of the segmentation is limited as well. We have analyzed a multiregional segmentation model based on the hybrid approach of the K-means algorithm which is driven by the ABC genetic algorithm. We suppose that the initialization distribution of the each cluster's centroid should reflect minimal variation towards the pixels lying inside the cluster. More the variation is increasing, worse results we obtain. Therefore, we define the fitness function minimizing the inter-cluster variance to obtain an optimal distribution of the image clusters within a predefined number of the ABC algorithm iterations. We have tested the segmentation procedure on a sample of the CT and MR image data, and verified this procedure against standard clustering algorithms.