Abstract
646
Objectives: There is an increasing interest in radiomic features calculation from PET images to characterize tumor heterogeneity. In this context, we are developing the concept of radiomic twins that goes beyond a standard classification approach. For each new patient, this concept consists in identifying a radiomic twin based on the patient tumor radiomic features in a patient database, ie another patient with a tumor presenting radiomic feature values similar to the ones observed in the new patient. The goal is to learn from the radiomic twin’s history and guide the patient management based on that history. This mimics the reasoning of a physician who uses his/her knowledge and experience with previous patients to select the best therapeutic option for a new patient. Here, we tested whether this concept could predict the tumor characteristics of breast cancer patients. Methods: 18F-FDG PET images from 138 patients at initial staging for a breast cancer were included. All patients were scanned using a Discovery 690 PET/CT scanner (GE Healthcare). In each patient, the primary lesion was segmented using a threshold set to 40% of SUVmax. In each resulting volume of interest, we computed 10 radiomic features using the LIFEx software (resampling step: 64 gray-levels between 0 and 20 SUV units) including SUVmax, SUVmean, Metabolic Volume (MV), TLG and 6 textural features. The process of radiomic twin identification included three steps. First, for each subject, we built a vector of 1 to 10 radiomic biomarkers, expressed as z-scores, called radiomic profile (RP). Second, we computed the Euclidian distance between the RP of the subject to be tested and the RP of each patient in the database. Finally, the smallest distance was used to identify the radiomic twin of the tested patient. We evaluated the ability of this approach to predict ten tumor characteristics (presence of metastasis, TNM stage, molecular subtype of cancer, Ki-67 expression level, HER-2 expression, lesion grade, presence of hormone receptors, necrosis, inflammatory stroma and in situ carcinoma) as a function of the vector composition by calculating the percentage of patients correctly classified. This procedure was repeated 50 times with 50% of patients randomly selected to produce the database and 50% of patients for whom the radiomic twin had to be identified. Results: The best performance was obtained when the RP included two radiomic features only, namely MV and Entropy, with 71±5% of the subjects having at least 6 tumor characteristics in common with his or her radiomic twin. With this 2-feature RP, 71±5% of patients had the same molecular subtype as the numerical twin, 69±5% the same status for the presence of in situ carcinoma, 69±4% the same HER-2 expression, 68±4% the same stage of necrosis, 67±5% the same status for the hormonal receptors, 66±4% the same Ki-67 expression level and 66±5% the same lesion grade. 66±5% of the subjects to be tested had both the same molecular subtype and the same hormone receptor status as their radiomic twin. Conclusions: We demonstrated that the radiomic twin approach identifies breast lesions with similar tumor characteristics. 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. Further studies involving more radiomic features and an independent cohort are still needed to better determine the potential of this approach for PET images. Additional investigations are also on-going by including clinical and biological data in the vector of features and by optimizing the method parameters, such as the distance definition and feature selection step to define the RP.