NON-SMALL CELL LUNG CANCER TUMOR CHARACTERISATION USING DEEP LEARNING

Abstract

Non-small cell lung cancer (NSCLC) constitutes the vast majority of lung cancers and has two major subtypes, adenocarcinoma (ADC) and squamous cell carcinoma (SqCC). Generally, these two subtypes are distinguished from each other by considering microscopic morphological criteria. However, poor morphology makes this quite difficult. Such studies are important for subspecialty treatment methods. In this thesis, deep learning (DL) methods on the subtype classification of NSCLC were investigated. In the first study, 73% success rate was achieved by using artificial neural networks (ANN), which form the basis of DL methods. In the second study, several DL models were investigated on subtype classification using segmented tumor slices from PET images. As a result, VGG16 and VGG19 emerged as the most successful models with a 95% F-score. Later, slice based studies were abandoned and patient based studies were initiated. In the third study, the use of three-dimensional (3D) data created by combining slices from each patient was not successful. In the fourth study, three different experiments were conducted in which PET images were directly used, cropped to include peritumoral areas, and segmented only tumor parts. This study demonstrated the positive effect of peritumoral areas and VGG19 reached an F-score of 74%. In the fifth study, transfer learning and fine tuning works did not yield successful results. The latest study involving CNN-based and ResNet-based shallow networks was promising with an Fscore of 71%.