Técnicas de reconocimiento automático de patrones aplicadas a imágenes hiperespectrales médicas

Abstract

Hyperspectral imaging is an emerging technology for medical diagnosis. Some previous studies have employed this technology for detecting cancer diseases. In this research work, a multidisciplinary team compounds by pathologists and engineers present a proof of concept of using hyperspectral imaging analysis in order to detect human brain tumour tissue inside pathological slides. The samples were acquired from four different patient diagnosed with brain cancer, specifically with high-grade gliomas. The hyperspectral capture system consists on a hyperspectral camera coupled with a microscope. This system works in the VNIR spectral range (from 400 nm to 1000 nm) with a spectral resolution of 3 nm. The images where then processed in order to remove the effect caused by the acquisition system. Later, and based on the diagnostic provided by pathologist, a spectral dataset containing only labelled spectra from normal and tumour tissue was created.

The data were then processed using three different supervised learning algorithms: Support Vector Machines, Artificial Neural Networks and Random Forests. The capabilities of discriminating between normal and tumour issue have been evaluated in three different scenarios, where the inter-patient variability of data was or not taken into account. The results achieved in this research study are promising, showing that it is possible to distinguish between normal and tumour tissue exclusively attending to the spectral signature of tissue.