A demonstration of the concept of numerical twins in esophageal cancer patients

Complete text of publication follows: Purpose: The characterization of tumor heterogeneity using textural features in radiomic analyses of PET images has shown promise to predict patient response or survival. In this context, the goal of this study is to identify for each patient a radiomic numerical twin who has similar radiomic feature values to learn from the numerical twin's history and guide patient management. Here, we test this concept to predict treatment response. Subjects and Methods: 107 patients with newly diagnosed esophageal cancer underwent pre-treatment ¹⁸F-FDG PET scan (data extracted from Ypsilantis et al, PLoS ONE 10(9):e0137036). All patients received a neoadjuvant chemotherapy and were later classified as non-responders (NR=69) or responders (R=38). In each patient, the primary lesion in the baseline scan was segmented using a threshold set to 40% of SUV_max. In each resulting volume of interest, 103 radiomic features were extracted including 85 textural or fractal features and 18 histogram indices. Each lesion was associated with a vector b of biomarkers. We computed the element-wise ratio between the vector b(p) of one patient p and the vector b(i) of each of the other 106 patients i (i=1,P-1). A patient N was identified as the radiomic numerical twin of patient p if the distance of b(p)/b(N) to 1 was the lowest among the P-1 distances. Its response to treatment was then predicted as the one observed in patient N. We evaluated the ability of this approach to predict treatment response when using 2 or 3 biomarkers in b by calculating the Youden index in a leave-one-out validation. We compared the results with logistic regression and support vector machine (SVM) models. Results: When including two biomarkers in b, the best performance using the numerical twin concept was obtained using Kurtosis and Energy with 87% NR lesions and 63.2% R lesions accurately classified (Youden=0.50). With three biomarkers (Kurtosis, Energy and Fractal Dimension mean), Youden index increased to 0.55. With 2 or 3 biomarkers, the logistic regression and SVM models always yielded Youden index less than 0.43. Conclusion: This concept of radiomic numerical twins is validated in esophageal cancer to predict treatment response. We found that lesions with similar radiomic profiles consisting of only 2 to 3 biomarkers had similar response to therapy. The identification of numerical twins could assist patient management in the future, based on the disease evolution in the patients used to identify the numerical twin