Disease Diagnosis using multimodal FTIR spectroscopy

Abstract

In this talk, I shall introduce the basis why FTIR spectroscopy is becoming a non-invasive optical tissue diagnosis tool and particularly show that clinical investigations related to malignancy and cancer detection by spectroscopic means have attracted attention both by the clinical and non-clinical researchers. The FTIR datasets are imported from the spectrometer into software written in-house in the Python or MATLAB environments. Data pre-processing is a very sensitive matter, with imposition of selection criteria to avoid pixels not covered by tissue and/or those that displayed excessively strong scattering effects. Spectra are usually vector normalized and phase corrected. In some cases, they are converted to second derivatives using a Savitzky–Golay smoothing filter, using a second order polynomial filter. Spectral datasets are subsequently converted to pseudocolor images using for instance hierarchical cluster analysis (HCA), which clusters patterns in a dataset based on their spectral similarity, and the most suitable method of clustering is chosen. Principal Component Analysis (PCA) is used for data exploration and dimensionality reduction for prediction models, thus decreasing training time and overfitting. Linear Discriminant Analysis (LDA), Partial Least Squares (PLS), Support Vector Machine (SVM) and Random Forest (RF) algorithms are commonly trained as classification methods, where their accuracy, sensitivity and specificity are assessed through cross validation tests. Examples on skin cancer, bone and breast cancer will be presented.

File with abstract

Speaker

Denise Zezell
Nuclear and Energy Research Institute IPEN-CNEN
Brazil

Discussion

Ask question