# IR Spectroscopy and Linear Support Vector Machine Analysis of Colorectal Liver Metastasis

## TL;DR

This paper uses infrared spectroscopy and machine learning to analyze liver metastasis from colorectal cancer and develop a faster cancer detection probe.

## Contribution

The study introduces a Decision Contribution Spectrum derived from IR spectroscopy to aid in tumor/nontumor classification and design a faster cancer probe.

## Key findings

- A Decision Contribution Spectrum was derived to represent average spectral contributions in tumor/nontumor classification.
- A linear SVM model was trained using Leave-One-Case-Out to avoid overtraining and enable simple feature selection.
- Results suggest potential for a fast mid-IR cancer probe with reduced wavenumber range and increased intervals.

## Abstract

The Colorectal Liver
Metastasis (CLM) library contains 756,096
full range Fourier transform infrared (FTIR) microscope imaging spectra
of 14 frozen tissue sections from 7 different consenting patients
with liver metastasis of colorectal origin. A subset of 30 windows
was defined in predominant tumor or nontumor regions for the training
or testing of linear support vector machine (SVM) models. Since the
number of consenting patients was small, the primary purpose of this
work was to establish a physical chemistry perspective on the infrared
(IR) spectroscopy of metastatic liver cancer using the large number
of available spectra. The linear SVM model was trained with a Leave-One-Case-Out
strategy to avoid case leakage, minimize overtraining, and offer simple
feature selection. The primary result is the derivation and measurement
of a spectral form for the average contribution to the tumor/nontumor
decision at each spectral step, i.e., the “Decision Contribution
Spectrum”. Finally, the results are used toward the design
of a fast mid-IR cancer probe for the operating room giving quick
tumor decisions despite a reduced range of wavenumbers and an increased
interval between wavenumber steps as compared to the FTIR input spectral
data.