Method for identification of spectral targets in discrete frequency infrared spectroscopy for clinical diagnostics.

Abstract

Spectroscopic imaging with discrete frequency infrared (DF-IR) has the potential to have a major impact on the clinical utility of IR imaging techniques for biochemical detection of disease. This can be achieved in real time using imaging at selected wavenumbers tuned to molecular absorptions of interest enabling tissue and disease-specific contrast to be obtained from the sample. However, selecting the appropriate wavenumbers to measure for DF-IR is critical, since vital diagnostic information could be missed. Here we demonstrate the application of partial least squares discriminant analysis (PLS-DA) and variable importance for projection (VIP) to identify key diagnostic wavenumber targets for the detection of dysplasia in human colon polyp sections. A small dataset, including 41 regions of interest (25 benign, 16 dysplastic; 5175 spectra in total), was selected from Fourier transform infrared (FT-IR) images of human colon polyp sections. PLS-DA was used to differentiate between benign and cancerous human colon polyp sections (sensitivity 95%, specificity 93% cross-validated), and VIP scores were calculated for all wavenumbers. A second PLS-DA model was then calculated using only variables that VIP identified as significant, reducing the number of wavenumbers to ~25% of the full dataset. The resulting cross-validated sensitivity and specificity (93 and 90%, respectively) indicate that the VIP method selects the key diagnostic wavenumbers for this dataset. Finally, a robust subset of variables was identified by selecting wavenumbers that exceeded the minimum VIP score in >95% of our validation iterations. A cross-validated PLS-DA model using only the robustly selected wavenumber targets (~20% of the original wavenumbers) resulted in sensitivity and specificity of 91 and 82%, respectively, indicating that PLS-DA and VIP are suitable approaches for the selection of wavenumber targets in DF-IR.