Survival Prognostic Modeling Using PET/CT Image Radiomics: The Quest for Optimal Approaches

Abstract

To develop radiomics models with the optimum performance at survival prognostication, two main challenges including the selection of imaging modality that reflects the most relevant information, and selection of optimum feature selection (FS) and machine learning (ML) algorithms must be addressed. In this study, we aimed to address both challenges simultaneously for survival prediction in head and neck cancer (HNC) using PET and CT modalities. A total number of 201 HNC patients from four different centres, adopted from The Cancer Imaging Archive (TCIA), were enrolled. Beside single-modality models, we developed multi-modality anatomo-functional models by fusing anatomical information (CT) and functional information ( <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">18</sup> F-FDG PET) at feature and image levels. Radiomics features were extracted with regards to the Image Biomarker Standardization Initiative and were harmonized via ComBat harmonization to eliminate the batch effect due to multiple centres. Then, combination of 6 different FS and 6 different ML methods were applied to all radiomics models (single-modality, feature- and image-level fusion models) to find the best combination for each. All FS and ML methods were able to cope with continuous time-to-event survival data, and were selected based on public availability. Multifactor ANOVA test was also performed to quantify the variance of the outcomes owing to the selection of radiomics model, FS, and ML method (p-value<0.05 was considered as threshold for statistical significance). The best performance was achieved by applying IBMA as FS and GBM as ML method on the PET model (C-index = 0.74), and applying IBMA as FS and glmnet as ML on image fusion WLS model (C-index = 0.73). However, optimum FS and ML method differed regarding the imaging modality in use. The selection of FS method generated the greatest proportion of total variance in models’ performances (18.3%), followed by the selection of imaging modality and ML method (6.6% and 3.7% respectively).