CSIRO PUBLISHING Rapid Communication www.publish.csiro.au/journals/ajc Aust. J. Chem. 2004, 57, 1139–1143 FourierTransform Infrared Imaging and Unsupervised Hierarchical Clustering Applied to Cervical Biopsies Keith R. Bambery, A Bayden R. Wood, A Michael A. Quinn, B and Don McNaughton A,C A Centre for Biospectroscopy and School of Chemistry, Monash University, Clayton VIC 3800, Australia. B Department of Obstetrics and Gynaecology, Royal Women’s Hospital, ParkvilleVIC 3052,Australia. C Corresponding author. Email: D.McNaughton@sci.monash.edu.au FTIR images of cervical tissue from patient biopsies were processed with an unsupervised hierarchical cluster- ing algorithm and compared with hematoxylin- and eosin-stained adjacent sections. Anatomical and potential histopathological features were clearly resolved in the resultant cluster maps. The mean extracted spectra assigned to each cluster indicate that the major spectral differences between the different cells in tissue predictably occur in the amide I region (1700–1570 cm −1 ) and the phosphodiester/glycogen region (1200–1000 cm −1 ). FTIR imaging in which a focal plane array mercury–cadmium–telluride detector and unsupervised hierarchical clustering is used shows potential as a rapid, non-subjective diagnostic tool in cervical pathology. Manuscript received: 25 May 2004. Final version: 9 September 2004. FTIR spectroscopy using purpose-built IR microscopes is now widely used in chemistry and many other disciplines to monitor chemical changes in small samples, but until recently the time involved in the collection of FTIR maps or images of samples has been too great to be thought useful. A new generation of infrared imaging spectrometers based on focal plane array (FPA) technology has changed this situation, and instruments capable of collecting large arrays of spectra in a short amount of time are now available. FTIR imaging spec- trometers in the mid-IR range can be equipped with focal plane arrays of mercury–cadmium–telluride (MCT) or InSb detectors that vary in size, and in which each pixel has a the- oretical spatial resolution of about 5.5 μm. Instruments such as the DigiLab ‘Stingray’ FTIR imaging spectrometer oper- ate in rapid scan mode and can be used to generate a mosaic of FTIR images that are subsequently stitched together to produce a multi-tiled image that enables the encapsulation of larger objects such as tissue sections. We have used IR microspectroscopy to investigate the chemical changes that occur upon the development of dis- eases such as cervical cancer, [1] with a view to develop- ing IR microspectroscopy as a diagnostic tool for cervical smears. As part of this general thrust, we recently recorded and analyzed IR maps of normal and abnormal cervi- cal tissue using single-point microscopy, and successfully correlated the resultant spectral images with pathological features. [1] FTIR image data can be processed in a univariate mode whereby chemical maps (also called functional group maps) based on peak intensity, peak area, or peak ratios can be rou- tinely generated with the software supplied with these types of instruments. While these methods can provide information on the distribution and relative concentration of a particular functional group, they are not very useful in terms of clas- sifying anatomical and histopathological features within the tissue matrix. Several authors have applied multivariate pat- tern recognition methods such as unsupervised hierarchical cluster analysis, [1–7] K-means clustering, [7,8] principal com- ponents analysis, [9] linear discriminant analysis, [10] artificial neural networks, [9] and fuzzy C-means clustering [7,11] to tis- sue analysis. These methods are aimed at classifying spectra based on similarity, and thus are used to discern anatomical and histopathological features based on underlying differ- ences in the macromolecular chemistry of the different cell and tissue types that constitute the sample. In this study, we applied FTIR–FPA imaging spectroscopy and unsupervised hierarchical clustering (UHC) to investi- gate the tissue architecture of the cervical epithelium. By selecting the spectral window for data processing, we showed that the combination of FPA technology and UHC can be a rapid, non-subjective analytical tool for the identification of anatomical features in cervical tissue. Cervical samples were obtained by cone biopsy at the Royal Women’s Hospital (Melbourne) from patients diag- nosed by cytology to have high-grade cervical dysplasia. The tissue samples were embedded in paraffin blocks and sliced by microtome into 4 μm tissue sections. Up to ten adjacent sections were separated into two groups. One group was mounted on a glass slide and stained with hematoxylin and eosin (H&E) for light microscope examination. The other group was deparaffinized and mounted on Ag/SnO 2 coated absorption/reflection slides for FTIR–FPA imaging © CSIRO 2004 10.1071/CH04137 0004-9425/04/121139