Bag-of-Words Algorithm to Predict Response to Treatment in Gist Tumors Using CT Scans

The aim of the study is to develop a mutation-based radiomic signature to predict the response to imatinib-based drug therapy in gastrointestinal stromal tumors. This study included 82 patients with GIST from two centers. Radiomic features were extracted from small regions inside the tumor on CT images to exploit tumor heterogeneity. We applied the Bag of Words algorithm (unsupervised approach) to local features to create new predictive signatures (BoWs). We trained different machine learning classifiers using both BoWs only (approach 1) and BoWs combined with global features extracted from the whole 3D (approach 2). The best model was obtained with the approach 2, reaching 91%, 63%, 87%, and 71% in terms of accuracy for sensitive patients, accuracy for non-responsive patients, precision for sensitive patients, and precision for non-responsive patients, respectively. The results highlight that the study of microscopic changes in heterogeneous tumors might enable to identify radiomics predictive signatures.