Radiomics features of 11[C]-MET PET/CT in primary brain tumors: preliminary results on grading discrimination using a machine learning model

1096 Background: Nowadays Artificial Intelligence (AI) algorithms have demonstrated remarkable progress in image-recognition tasks and could be useful for the new concept of precision medicine in order to help physicians in the choice of therapeutic strategies for brain tumors. Previous data suggest that, in primary brain tumor, amino acid PET may demarcate more accurately the active disease than paramagnetic enhanced MRI that is the standard method of the evaluation in brain tumors and help in the evaluation of disease grading, as a fundamental basis for the correct clinical management of these patients. The aim of this study is to propose a radiomics workflow in order to create a predictive model on 11[C]-MET PET/CT scan images capable to discriminate between low-grade and high-grade primary brain tumors. Materials and Methods: in this retrospective study we selected fifty-six patients affected by primary brain tumor who underwent 11[C]-MET PET/CT from January 2016 to December 2019 at the Nuclear Medicine department of the Cannizzaro Hospital of Catania. Histological specimen the analysis was available in all patients in order to confirm the diagnosis and grading disease. PET/CT acquisition was performed after 10 minutes from the administration of 11C-Methionine (401-610 MBq) for a time acquisition of 15 minutes. 11[C]-MET PET/CT images were acquired using two scanners (24 patients on a Siemens scan and 32 patients on a GE scan). Then, LIFEx software was used to perform the segmentation and the extraction of features in order to identify the volume of interest (VOI) and to extract features of first, second, and third-order. LIFEx calculates 44 RF reflecting the VOI shape, the VOI voxel values, the histogram of VOI values, or the VOI textural content. A novel mixed descriptive-inferential sequential approach was used to identify a subset of relevant features that correlate with the outcome. A machine learning model based on discriminant analysis was tested in the evaluation of grading prediction (low vs high-grade) of 11[C]-MET PET/CT. Results: The machine learning model applicated in this study to stratify brain tumors as “low” and “high” grade showed discriminant texture features as follow: two features in fixed ROI respectively NGLDM Busyness (p-value: 0.1615) and GLZLM LZLGE (p-value 0.3207); three features in total volume respectively SHAPE Sphericity (p-value 0.0314), SHAPE Compacity (p-value 0.0215) and HISTO Kurtosis (p-value 0.0232). The results of performances, using a semi-automatic and user-independent segmentation process and an innovative feature extraction process on fixed ROI, showed a good performance in prediction of disease grade with an accuracy of 70.31% (AUC 64.13%) in all patients, of 80.51% (AUC 65.73%) in GE tomograph, of 84.98%(AUC 78.91%) in Siemens Tomograph. Using our model of machine learning with discriminant analysis the results of the performances were suboptimal in all patients group( 24 Siemens plus 32 GE) with an accuracy of 57.25 % and AUC of 58.51%. However in subgroup analysis, despite the low number of patients and different scanner, our model demonstrated good performances in the prediction of disease grade as follow: GE Scanner Accuracy 71.64 and AUC 62.8%; Siemens Scanner Accuracy 72,88% and AUC 78.91%. Conclusions: This preliminary study on radiomics features analysis of 11[C]-MET PET/CT suggests a possible value of Artificial Intelligence to assess primary brain tumors and to predict disease outcome in terms of grading discrimination at diagnosis. Further studies are needed to improve radiomics algorithms to personalize predictive and prognostic models and potentially support the medical decision process.