Work

P921: Artificial Intelligence Based Fdg Pet/ct Radiomics for Risk Stratification in Newly Diagnosed Multiple Myeloma

Year: 2023

Type: article

Abstract: Topic: 14. Myeloma and other monoclonal gammopathies - Clinical Background: Multiple myeloma is characterized by substantial biological heterogeneity, which is reflected in highly variable survival outcomes and represents a challenge for risk stratification. The established prognostic models rely on serum biomarkers (ISS) and the incorporation of bone marrow biomarkers (R-ISS, R2-ISS) has led to modest performance improvements. Imaging findings are of prognostic significance but have not yet been incorporated into risk stratification models. Aims: To extract radiomics features from FDG PET/CT scans of newly diagnosed patients using a deep learning model and to develop an integrated risk stratification model (peripheral blood, bone marrow, radiomics). Methods: After pre-processing of the native DICOM images, we extracted discrete radiomics features from the diagnostic FDG PET/CT scans of patients with newly diagnosed multiple myeloma using the nnU-Net architecture (with default parameters, 41% adaptive SUVmax thresholding) with a model trained on external patients with hematologic malignancies and PyRadiomics. Limitations of this approach include heterogeneity introduced by varying imaging quality, technical differences related to PET reconstructions, and potential inclusion of non-neoplastic FDG activity in segmented regions. We evaluated the prognostic significance of 10 select tumor shape- and texture-related radiomics features. We integrated these radiomics features with demographic information, peripheral blood biomarkers, and bone marrow cytogenetics. For modeling purposes, patients without measurable metabolic tumor volume were categorized as having undetectable radiomics features (fixed at value 0 for all features). All radiomics features were then divided into quintiles for analysis (dichotomizing the data at the most extreme quintile). For comparative modeling, we partitioned the patient population into 50% training and 50% test subsets. The outcome of interest was overall survival and model performance was assessed by comparing Harrell’s C-statistic among competing models. Results: We analyzed the diagnostic FDG PET/CT scans of 748 patients with newly diagnosed multiple myeloma seen at Mayo Clinic between 2004 and 2019. Median age at diagnosis of these patients was 64 years (29-94) and 460 of them (62%) were men. After a median follow-up of 1.8 years (95% CI 0.1-7.4 years), median overall survival was 7.9 years (95% 6.3-NR). Most patients had measurable metabolic tumor volume (n=505, 68%) on their diagnostic FDG PET/CT. The mere presence (compared to the absence) of measurable metabolic tumor volume was not a prognostic factor (HR 0.99, 95% CI 0.71-1.38, p=0.938), but measures of tumor shape and texture were. In the entire cohort, radiomics features related to tumor texture (top 20% variance in gray level intensities, a measure of tumor homogeneity across volume, GLV, Figure A) and shape (bottom 20% surface-to-volume ratio, SVR, Figure B) were strongly associated with overall survival, independent of ISS, R-ISS, and each other. Five hundred and fifty-five patients had complete ISS and 443 had complete R-ISS data, allowing for model comparison. The addition of the two features to a risk stratification model containing age and ISS led to an improvement in model discrimination (C-statistic +1.4%, 95% CI 0.2-2.6, p = 0.029). Conclusion: Radiomics features obtained from diagnostic FDG PET/CT scans of multiple myeloma patients using a deep learning model were strongly associated with overall survival. This association was independent of age and the established ISS models and the radiomics features were able to further improve existing risk stratification models. Standardized radiomics analysis represents a promising approach to improve risk stratification by non-invasive means.Keywords: FDG-PET, Multiple myeloma, Prognostic factor, Artificial intelligence

Source: HemaSphere

DOI: 10.1097/01.hs9.0000970588.15601.1e