From preprocessing to fuzzy classification of IR images of paraffin embedded cancerous skin samples

Abstract

Mid-Infrared (IR) micro-spectral imaging is an efficient method to analyze molecular composition of biomedical samples. In clinical oncology, this non-invasive technique is generally used on frozen biopsies to localize and diagnose cancerous tissues in their early stages. However, samples are usually fixed in paraffin in order to be preserved from decay, but the IR signature of paraffin prevents the study of the underlying tissue. To neutralize the paraffin signal from the recorded data, preprocessing methods based on Independent Component Analysis (ICA) and Nonnegatively Constrained Least Squares (NCLS) or on Extended Multiplicative Signal Correction (EMSC) have been recently developed. Then, in order to identify tumor areas, clustering techniques are applied on the preprocessed data, the final result being a false-color map of the biomedical sample which is comparable to the conventional histological image. By allowing each recorded spectrum to be assigned to every cluster, the fuzzy clustering gives more realistic results for unclear tissue boundaries by better highlighting the tumor and peritumoral areas. A recent algorithm based on the redundancy of classes allows to automatically estimate the optimal number of classes and the optimal fuzzy parameter. In this paper, we analyze the effects of the preprocessing methods on the optimal parameter extraction and on the results of the fuzzy clustering on different paraffin embedded cancerous skin samples.