Body composition radiomic features as a predictor of survival in patients with non-small cellular lung carcinoma: A multicenter retrospective study

Abstract

ObjectivesThis study combined two novel approaches in oncology patient outcome predictions—body composition and radiomic features analysis. The aim of this study was to validate whether automatically extracted muscle and adipose tissue radiomic features could be used as a predictor of survival in patients with non-small cell lung cancer. MethodsThe study included 178 patients with non-small cell lung cancer receiving concurrent platinum-based chemoradiotherapy. Abdominal imaging was conducted as a part of whole-body positron emission tomography/computed tomography performed before therapy. Methods used included automated assessment of the volume of interest using densely connected convolutional network classification model - DenseNet121, automated muscle and adipose tissue segmentation using U-net architecture implemented in nnUnet framework, and radiomic features extraction. Acquired body composition radiomic features and clinical data were used for overall and 1-y survival prediction using machine learning classification algorithms. ResultsThe volume of interest detection model achieved the following metric scores: 0.98 accuracy, 0.89 precision, 0.96 recall, and 0.92 F1 score. Automated segmentation achieved a median dice coefficient >0.99 in all segmented regions. We extracted 330 body composition radiomic features for every patient. For overall survival prediction using clinical and radiomic data, the best-performing feature selection and prediction method achieved areas under the curve–receiver operating characteristic (AUC-ROC) of 0.73 (P < 0.05); for 1-y survival prediction AUC-ROC was 0.74 (P < 0.05). ConclusionAutomatically extracted muscle and adipose tissue radiomic features could be used as a predictor of survival in patients with non-small cell lung cancer.