Spatial sampling effect on data structure and random forest classification of tissue types in High Definition and Standard Definition FT-IR imaging

Abstract

Infrared radiation imaging (IR) combined with machine learning algorithms is one of the most promising techniques in tissue structures recognition and cancer diagnosis. Pancreatic tissue is heterogeneous and provides variable biochemical composition, as well as exhibits a variety of shapes and sizes having strong impact on physical aspects of light absorption. This research explores the potential of High and Standard Definitions in pancreatic tissue type prediction with random forest classification based on different spectral ranges. IR images spatial resolution depends on wavenumber and objective's numerical aperture. Therefore, data measured in Standard Definition (SD) are affected by under-sampling in the high wavenumbers while High Definition (HD) is free of this limitation for the whole analyzed spectral region. In order to investigate this effect a Random Forest model was created using a set of 56 biopsies measured in both definitions. In terms of signal variance the HD was found to be more spread out (however, with smaller scattering). SD models obtained a higher True Positive Rate than HD, however, spatial detail was better in HD models. Models trained on one definition and predicted on the other suggest that HD model can be used to successfully predict SD data. Finally, models based on only fingerprint and only high wavenumber were created to assess information content in each region. Models based on fingerprint region had very good prediction results across all classes, while for high wavenumber region results were class and definition dependent.