On Stability of Feature Selection Based on MALDI Mass Spectrometry Imaging Data and Simulated Biopsy

Abstract

In this work we analyse MALDI mass spectrometry imaging data for thyroid cancer samples. Such a data, containing information about spatial distribution of proteins/peptides, makes possible to make a virtual analysis how a technique of fine needle aspiration (FNA) biopsy, a routine diagnosis procedure for thyroid, influences the outcome i.e. a set of discriminative features between cancerous and normal tissue. We hypothesised that an impure dataset (consisting of normal cell contaminated cancer samples) would be beneficial in the terms of stable feature selection. We compared several methods of predictor selection on different datasets to perform an in-depth feature ranking stability analysis for thyroid cancer mass spectrometry data. Furthermore we examined the impact of sample contamination level on the selection.