Linear Support Vector Machine Model of an IR Spectral Library: Application to Skin Cancer

The primary result of this work is the derivation and measurement of the average "Decision Contribution Spectrum" using a mouse tumor sample data set and the linear support vector machine (SVM) method from a perspective of spectroscopy. The "Decision Contribution Spectrum" gives the average contribution to the decision (tumor/nontumor in this case) at each step along the spectrum. A library of more than four thousand infrared (IR) spectra was obtained with a Fourier Transform Infrared (FTIR) microscope imaging system on a frozen section of an SKH1 mouse tumor - an accepted murine model for studying squamous cell carcinoma in that it very closely recapitulates the human disease. A linear SVM model was trained and tested avoiding overtraining and offering simple feature selection. It was used to see how much data can be removed without affecting the quality of decisions. Then, two further efforts are described to move IR spectroscopy toward future use in human skin cancer detection: (i) the design of a reduced range and reduced sampling for a fast and hand-held, mid-infrared spectral probe, and (ii) the use of a fiber-loop sensor probe (with FTIR in this preliminary study) on live SKH1 mice that had tumors to detect cancer externally and show no ill effects on the mice. The combination of the latter efforts supports the feasibility of using a fiber-loop sensor with a fast and hand-held mid-infrared spectral probe for the detection of skin cancer on humans. This work does not demonstrate a working human skin cancer probe, rather it provides evidence for judging whether work on live human skin is justified.