Investigating the radiomic signatures of metastatic and non-metastatic breast cancer patients based on F18-FDG PET/CT images.

641 Objectives: Radiomic data mining has been an expanding research field for the past ten years. This promising domain could provide new diagnostic and predictive models to assist physician decision making. In this work, we investigate whether radiomic features calculated from FDG PET images in breast cancer patients were associated with the metastatic status of the patients. Methods: Pre-treatment 18F-FDG PET images from 161 patients with a breast cancer were included. Patients received an intravenous injection of 3-3.5 MBq/kg of FDG when capillary blood glucose level was below 11 mmol/L and after 6 hours fasting. All patients were scanned ~60 min post-injection using a Discovery 690 PET/CT scanner (GE Healthcare). Images were reconstructed using an ordered-subset expectation maximization reconstruction algorithm (2 iterations, 24 subsets) and post-filtered using a 6.0 mm full width at half maximum Gaussian filter, with a voxel size of 2.7x2.7x3.3 mm3. In each patient, the primary lesion was segmented using a threshold set to 40% of SUVmax. For each tumor, 42 radiomic features were calculated from the tumor region using the LIFEx software (resampling step: 64 gray-levels between 0 and 20 SUV units) including SUVmax, SUVmean, Metabolic Volume (MV), Total Lesion Glycolysis (TLG) and textural features. Each tumor was thus characterized by a radiomic profile (RP) of 42 values. The RP were randomly separated in a training set of 121 RP and a testing set of 40 RP. The training set RP were converted in z-scores and then divided into an M+ group of 23 metastatic patient RP and an M- group of 98 non-metastatic patient RP. Non supervised RP clustering was used in the M+ and M- groups to identify a small number of M+ and M- clusters and the mean RP of each cluster was used to define a signature for each cluster. The significance of the difference of each feature value between the M+ and M- signatures was used to select a subset of most discriminant features hence reduce the RP dimension of each signature. To determine the usefulness of these signatures, each of the 40 RP expressed as z-scores of the testing set was assigned to the M+ or M- category based on the shortest Euclidian distance to a M+ or M- signature. Results:In the M+ group, 6 signatures were identified, suggesting that M+ tumors have a variety of radiomic profiles. Similarly, 11 signatures were found in the M- group, demonstrating high radiomic heterogeneity within the M- tumors. 13 radiomic features (Volume, TLG, Homogeneity, Energy, LRE, LGRE, SRLGE, LRLGE, RLNU, LZE, LGZE, SZLGE and LZLGE) were significantly different between the M+ and M- signatures and the reduced RP therefore included 13 values. When associating the 40 RP of the testing set to M+ and M- signatures, only 3 of the 8 (38%) metastatic subjects and 22 of the 32 (69%) non-metastatic subjects were properly classified, showing that the M+ and M- radiomic signatures were not very specific (Figure 1). Conclusions: Our results suggest that the radiomic features of BC tumors as measured on the baseline PET are not closely related to the metastatic status of the patients. M+ patients can present with various radiomic profiles, and so can M- patients. In addition, radiomic profiles observed in the M+ patients are similar to those observed in the M- group, so that the radiomic profile alone does not help identify the metastatic status of the patient. Further studies are needed to determine whether associating other features to the radiomic profile could give useful information regarding the metastatic status of the disease.