This work serves as a preliminary study to investigate and identify the applicability of domain knowledge as an initialization method for K-Mean (KM), typically in medical images. For this purpose, 20 breast histopathology images were used as data set and the evaluations are focused on the clustering of the hyperchromatic nucleus. The iteration numbers and clustering results (i.e., accuracy, over-segmentation, and under-segmentation) are benchmarked with KM++ and the conventional random initialization method. The domain knowledge initialization method is found promising by achieving lower iteration numbers (<9), higher percentage in accuracy (85.5% (±2.27)), and lower percentages in over-segmentation (8.25% (±2.23)), and under-segmentation (7.00% (±2.14)). From this study, we hypothesize that the domain knowledge initialization method has the potential to be implemented as an initialization method and is posited to overperform some established initialization methods, typically for clustering tasks in medical images.
