TY - JOUR T1 - <strong>Towards the identification of tumors with similar radiomic features: introducing the concept of numerical twins</strong> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 502 LP - 502 VL - 58 IS - supplement 1 AU - Fanny Orlhac AU - Christophe Nioche AU - Ingrid Faouzi AU - Michael Soussan AU - Irene Buvat Y1 - 2017/05/01 UR - http://jnm.snmjournals.org/content/58/supplement_1/502.abstract N2 - 502Objectives: The characterization of tumor heterogeneity using textural features (TF) in PET has shown promise to predict patient response or survival and is the basis of radiomics. In this context, our goal is to identify a set of radiomic features that would be specific of a tumor type and then use that model to identify so-called numerical twins that would present similar radiomic feature values and hopefully belong to the same tumor type (eg, same histological features). Here, we present a first demonstration of that concept.Methods: 17 patients with non-small cell lung cancer (NSCLC) underwent a PET scan 115 minĀ±15 after injection FDG on a Sigma PET/MR machine. In each patient, the primary lesion was segmented using a threshold equal to 40% of SUVmax. In each resulting volume of interest (VOI) from PET images, we computed SUVmax, SUVmean, metabolic volume (MV), Total Lesion Glycolysis (TLG) and 6 TF using the LIFEx software (VOI quantization between 0 and 40 SUV units using 128 gray-levels): Homogeneity, Entropy, SRE, LRE, LGZE and HGZE. Each lesion k was thus associated with a vector b of 10 biomarkers b(k,j) (4 conventional indices and 6 TF). The vector most representative of adenocarcinoma (ADK) tumors (vector m1 composed of biomarkers m1(j)) and the one most representative of squamous cell carcinoma (SCC) tumors (vector m2 composed of biomarkers m2(j)) were identified. The similarity S1 of any lesion k with vector m1 was computed as the standard deviation across the 10 ratios b(k,j) /m1(j) (and same for S2). This metric is 0 if b(k,j)=A.m1(j) for all j biomarkers, where A is a scalar (i.e. b(k) and m1 are homothetic). A lesion was identified as numerical twin of m1 if S1&lt;S2 and of m2 if S2&lt;S1.Results: 12 lesions were adenocarcinomas and 5 were squamous cell carcinomas. Eleven lesions were identified as numerical twins of the ADK model and 4 as numerical twins of the SCC. Among the 11 ADK twins, 9 were histology-proven ADK, while 2 out of the 4 SCC twins were histology-proven SCC. Overall, 73% of the lesion were accurately classified.Conclusion: This concept of numerical twins defined by a metric S supporting the comparison of vectors including biomarkers of different units is introduced. We found that lesions with similar radiomic profiles consisting of only a few biomarkers had pathologic similarities. Additional NSCLC patient scans are currently being included to validate the concept on a large cohort and MR biomarkers will also be included. The identification of numerical twins could assist patient management in the future, based on the disease evolution in the patients used to define the model vectors. Research Support: ANR-11-IDEX-0003-02 ER -