Understanding omics data of lung cancer patients: Correlations between metabolomics and radiomics

Abstract

Eur J Nucl Med Mol Imaging (2021) 48 (Suppl 1): S1-S648 predicting response and survival was obtained by combining clinical data with PET and CT texture parameters (AUC 0.87). Conclusion: PET/CT derived parameters demonstrated better performances-than the clinical parameters in predicting the response and overall survival confirming the interest in considering a radiomics based approach for the optimization of therapy management in patients with head and neck cancer. References: none Aim/Introduction: Qualitative and semi-quantitative parameters of PET and CT images are used to assist decision making for cancer treatment. In initial studies PET and CT radiomic features have shown promising results in disease prognostication and treatment outcome prediction in cancer. These features are specific to the outcome and different features show association with different outcomes. Hence finding the scalability of radiomic features from one modality to another can have promising impact. In our study we have tried to check the scalability of radiomic features across the modalities, PET and CT. We have performed a study to predict CT radiomic features using PET radiomic features and vice versa. Materials and Methods: This study was approved by the institutional ethics committee as retrospective study. 104 NSCLC patients who underwent pre-treatment PET-CT scan were included in this study. Primary lung tumor was delineated by SUV cutoff (42%) method on PET images and saved as RTStructure for PET and CT series. These Images and RTStructures were used for radiomic extraction using bin-width of 25 and 5 for CT and PET respectively using pyradiomic 2.1.0 software and in-house developed python script. Subsequently, concordance correlation coefficient (CCC) was calculated between PET and CT features and top 25 correlated features (excluding shape features) were selected to develop a prediction model. Entire set of data was split into training and validation sets (70:30). For each PET radiomic feature; a set of CT features were selected and vice versa using Recursive Feature Elimination(RFE). For individual feature prediction across modalities, a multivariate linear regression model was developed using selected features. Model performance was assessed based on accuracy of prediction (C-index) on validation set. Results: Around 54% and 46 % radiomic features show positive and negative correlation across PET and CT respectively. Only 91(8.33%), 69(6.3%) and 51(4.67%) features have 0.5<CCC<0.7, 0.7≤CCC<0.9 and CCC≥0.9 respectively. Top 25 selected radiomic features had CCC equal to or more than 0.99. The average C-Index and p-value in validation set for 25 PET radiomic features prediction was found to be 0.988(±0.019) and <0.0001 respectively. Similarly, average C-Index value and p-value in validation set for 25 CT radiomic features prediction was found to be 0.987(±0.016) and <0.0001 respectively. Conclusion: As per our findings, very few radiomic features have good correlation between PET and CT. These features show excellent capability to predict features across these modalities. References: none Aim/Introduction: Treatment of lung cancer remains challenging, partly due to the late-stage diagnosis of patients. With a strong focus on non-small cell lung cancer (NSCLC), this pilot study examines the diagnostic and prognostic potential of combining specific metabolic biomarkers from blood plasma (metabolomics) with features out of medical images (radiomics). This way, metabolomics and radiomics might be at the base of developing a more personalized treatment plan for lung cancer patients. This study aims to combine a metabolomics and radiomics dataset from lung cancer patients and to unravel the underlying correlations between the techniques. Materials and Methods: The initial patient cohort consisted of 32 patients, all diagnosed with early-stage NSCLC. All patients underwent a lobectomy as part of their standard-of-care treatment plan. The PET-CT images of all the patients were collected using 18 F-FDG (Biograph Horizon camera, Siemens). The PET-CT images were then segmented using a semi-automatic tool (ACCURATE), creating specific volumes of interest (VOIs) of the lung lesions for each patient. By loading the VOIs into the second tool (RADIOMICS), 486 radiomics parameters were extracted from each VOI (Both tools developed by R.B.) Simultaneously, 238 metabolic parameters representing 62 plasma metabolites were determined from the same patients using proton nuclear magnetic resonance (1 H-NMR) spectroscopy. A correlation coefficient test was used on the total omics-dataset to find a correlation between these two sets of parameters. Results: The correlation values found between the radiomics and metabolomics parameters showed R 2 values between 0.5 and 0.65 (positive correlation) or between-0.5 and-0.65 (negative correlation). The positive correlations found in the metabolomics dataset were mainly related to the concentration of plasma glucose. The radiomics S509