A Comparison of Fuzzy and Non-fuzzy Clustering Techniques in Cancer Diagnosis

In this paper, we apply K-means and Fuzzy C-Means, two widely used clustering algorithms, to cluster a lymph node tissue section which had been diagnosed with metastatic infiltration (cancer spread from its original location). Each cluster algorithm is run 10 times as different initialisation states may lead to different clustering results. We compare the performance of the two algorithms by subjectively altering the number of clusters from 2 to 9 and we analyse the results using false-colour images (which are produced as a function of the spatial coordinates on the tissue section). In the initial stages of this experiment, we found that the ranges of the first three principal components were too small and may lead to small objective function values in Fuzzy C-Means. Therefore, the minimal amount of improvement must be set to a small enough value to allow the cluster centre positions to improve, otherwise the iteration will stop prematurely. After adjusting this setting, the performance of Fuzzy C-Means was significantly better. The results show that Fuzzy CMeans can separate the major different tissue types using just a small number of clusters, whereas K-means is only able to separate them if a larger cluster number is used.