The Application of a Simulated Annealing Fuzzy Clustering Algorithm for Cancer Diagnosis

Fourier Transform Infrared Spectroscopy (FTIR) is becoming a powerful tool for use in the study of biomedical conditions, including cancer diagnosis. As part of an ongoing programme of research into the potential early diagnosis of cervical cancer, Hierarchical Cluster Analysis (HCA) and Fuzzy C-Means (FCM) have been applied to distinguish FTIR spectra obtained from cancerous and non-cancerous cells. In recent experimentation on non pre-processed FTIR spectra data, the FCM method has been shown to achieve significantly better results. Nevertheless, two limitations apply to this technique. Firstly, a priori assumption of the number of clusters is needed. This is a problem because, in general, this is not known in medical diagnosis. The other limitation is that it involves a greedy local search methodology such that sub-optimal solutions may be returned and, thus, misdiagnosis could occur. Bandyopadhyay [8] has recently proposed a Simulated Annealing Fuzzy Clustering algorithm (SAFC) which can avoid these two limitations. However, when we implemented the proposed algorithm, it was found that sub-optimal solutions could be obtained in certain circumstances. In this paper, we extend the SAFC algorithm to overcome this difficulty and apply this modified version to the classification of seven sets of FTIR spectra data which have been taken from three oral cancer patients. With no prior specification of cluster number, our modified SAFC algorithm is shown to obtain the correct (clinical) classification of clusters in 4 out of 7 data sets. In the remaining 3 data sets it produces a number of clusters which, while differing from the clinical classification, appear to better match the underlying data when subjectively visualised using Principal Component Analysis (PCA).