Machine learning applied to imaging: validation of the radiomics approach in a population of non-small-cell lung cancer patients treated by (chemo-)radiotherapy at the Cliniques universitaires Saint-Luc

Madeleine Scrivener Published in the journal : November 2018 Category : Mémoires de Recherche Clinique

Summary :


An increasing number of advances have been achieved in the field of medical imaging, namely the conversion of standardof- care images into mineable data. Radiomics refers to the high-throughput extraction of quantitative image features from medical images (watch the video on Tq980GEVP0Y and visit These image features can be divided into four groups depending on the tumor characteristic they describe: tumor intensity, tumor shape, tumor texture, or wavelets. This study consisted in a radiomics analysis of 18’026 features extracted from standard-of-care, pretreatment, 4D computed tomography (CT) images from a cohort of 44 non-small-cell lung cancer (NSCLC) patients treated with chemoradiotherapy at the Cliniques Universitaires St-Luc. A radiomics signature was created using machine learning algorithms in order to preselect a small group of radiomics features based on their correlation with the studied endpoint (survival, histological types, etc.). The signature was created by analyzing the features of a patient cohort, and its robustness was then tested and validated on an external independent dataset.



First, we hypothesized that adenocarcinomas could be differentiated from squamous cell carcinomas using a new radiomics signature. This would show that the radiomics features extracted from pretreatment CTs contain information about the tumor’s histological type. We further hypothesized that we could validate two previously published prognostic radiomics signatures (Aerts et al. and Tunali et al.) on the St-Luc cohort.



Our study was divided into four parts. In the first part, we created a radiomics signature in order to differentiate histologically confirmed adenocarcinoma from other histological types in an open-source cohort of 422 NSCLC patients from the Maastricht University. We then validated the signature in the cohort of 44 patients treated at St-Luc and in another cohort of 99 NSCLC patients treated in the Department of Radiation Oncology at the University of California, San Francisco (UCSF). In the second part, we validated the Aerts et al. prognostic signature published in 2013 in Nature Communications in the St-Luc cohort. This signature uses four radiomics features describing the tumor so as to divide the patients into two different prognostic groups. Thirdly, we validated the published signature of Tunali et al. in the St-Luc cohort. This signature is also a prognostic signature that uses only two radiomics features, namely radial gradient (RG) and radial deviation (RD), to divide the patients into a group with better prognosis and more indolent tumor and another group with more aggressive disease and poorer prognosis. Lastly, we used two different international scoring systems to evaluate the methodology of our approach: the Radiomics Quality Score (RQS) and the TRIPOD recommendations.



The first part of the study showed significant results for training (AUC: 0.93) and validation (AUC: 0.82) of the histological radiomics signature. The two prognostic radiomics signatures showed promising, though not statistically significant, results, thereby highlighting the potential of radiomic features and machine learning for non-invasively providing additional information on the tumor’s proliferation and histological type. This could eventually be used in routine practice, thus improving therapeutic decision-making and reducing complications associated with invasive diagnostic procedures. This study will be continued on a larger cohort of St-Luc patients.