FCM parameter estimation methods: Application to infrared spectral histology of human skin cancers

Spectral histology of cancer can be achieved thanks to the analysis of infrared (IR) hyperspectral images. The Fuzzy C-Means (FCM) clustering is particularly well adapted since each object is attributed to all clusters with different membership values. Applied on IR hyperspectral images of human skin cancers, it can highlight fine transitions between tumor and surrounding tissues and/or tumor heterogeneities. However, to provide a biomedically interpretable clustering, the relevant values of the two FCM parameters, i.e. the number of clusters and the fuzziness parameter, must be judiciously selected for each analyzed tissue. In this paper, the performance of some classical cluster validity indices and m-rules previously presented in the literature are evaluated. A new heuristic method based on the redundancy of FCM clusters is also proposed. We show that our method highly improves the clustering quality and computational time when applied on IR images of human skin tumors.