Mathematical morphology applied to Radiomics

Details

Supervisors

Macq, Benoît ; De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Personalized medicine is a major and increasing challenge, especially in the diagnostic and treatment of cancer. Traditionally medical imagery was used, but often in a qualitative and subjective way. Radiomics proposes to combine quantitative image features with those extracted from other techniques (clinical, genomic...) to help cancer diagnostic and prediction. The objective of this thesis is to develop new mathematical morphology-based features, and use them for the first time with Radiomics. We started with listing and analyzing the mathematical foundations of mathematical morphology and the concept of morphological series in order to extract scalar features. These were then compared with classical Radiomic features, namely first order features and gray level matrix based texture descriptors. These features were implemented and experimented on an open database of 3D Xray-CT segmented non-small cells lung cancer (NSCLC) tumours. A Kaplan-Meier survival analysis was then used in order to select the features that are statistically proven to be relevant. In addition a different machine learning decision tree approach is also tried and the feature selection of the algorithm is compared to the results of the survival analysis. The comparison of our new mathematical morphology-based features to the classical Radiomics features managed to show that the mathematical morphology-based features are at least as good as the classical features extraction techniques. Our contribution brings to the table new image bio-markers based on mathematical morphology that are statistically proven effective on this database, and could be used for future Radiomics research and personalized medicine.