Intelligent Measurement Analysis on Single Cell Raman Images for the Diagnosis of Follicular Thyroid Carcinoma

Inter-observer variability and cancer over-diagnosis are emerging clinical problems, and there is a strong necessity to support the standards histological and cytological evaluations by biochemical composition information. Over the past decades, there has been a very active research in the development of Raman spectroscopy techniques for oncological applications and large scale clinical diagnosis. A major issue that has received a lot of attention in the Raman literature is the fact that variations in instrumental responses and intrinsic spectral backgrounds over different days of measurements or devices creates strong inconsistency of Raman intensity spectra over the various experimental condition, thus making the use of Raman spectroscopy on a large scale and reproductive basis difficult. We explore different methods to tackle this inconsistency and show that regular preprocessing methods such as baseline correction, normalization or wavelet transformation are inefficient on our datasets. We find that subtracting the mean background spectrum estimated by identifying non-cell regions in Raman images makes the data more consistent. As a proof of concept, we employ our single-cell Raman Imaging protocol to diagnosis challenging follicular lesions, that is known to be particularly difficult due to the lack of obvious morphological and cytological criteria for malignancy. We explore dimensionality reduction with both PCA and feature selection methods, and classification is then performed at the single cell level with standard classifiers such as k Nearest Neighbors or Random Forest. We investigate Raman hyperspectral images from FTC133, RO82W-1 and NthyOri 3-1 cell lines and show that the chemical information for the diagnosis is mostly contained in the cytoplasm. We also reveal some important wavenumber for malignancy, that can be associated mainly to lipids, cytochrome and phenylalanine.