Texture Analysis of Histopathology Slides for the prediction of EGFR Gene Mutation

Abstract

This thesis studies the effect that a somatic mutation has on the texture format of a tumor histopathol- ogy slide. We examine specifically the case of the EGFR mutation on Adenocarcinoma Lung Cancer type, with use of the latest computational methods. Adenocarcinoma Lung Cancer accounts for about 40% of all lung cancers and during the latest years, many targeted therapies related to specific gene mutations, such as the EGFR gene, have been developed. The link between the mutation and the histopathology slide can lead to faster, more accurate diagnosis as well as valuable pattern detection. The problem is firstly addressed and analyzed both biologically and computationally, to deter- mine the best possible approach. Following this, two different computational methods are developed with the purpose of detecting texture feature within tissue slides characterized by an EGFR mutation. The first method used is Convolutional Neural Networks for image recognition and achieves a good classification rate. Further analysis for the origin of the produced features is needed though, and there- fore, the second method of Digital Image Texture Analysis is applied. That achieves an even better success rate, which strongly implies the existence of texture features connected to the EGFR mutation. At the same time, LUAD gene expression subtypes, are also explored in terms of texture dif- ferences on their corresponding tissue slides. This is performed mainly to distinguish the features detected between a specific subtype, TRU, which is enriched with EGFR mutation, and EGFR muta- tion related texture features. Finally, a thorough presentation of the above results is made, which all conclude that there is a distinct connection between the presence of an EGFR somatic mutation and its effects on the texture appearance of the tissue slide.