Circulating plasma gelsolin and MRI-based radiomics as biomarkers of platinum resistance in epithelial ovarian cancer: building a multiparameteric prediction algorithm.

backgroundResistance to platinum-based chemotherapy in epithelial ovarian cancer (EOC) patients is a barrier to disease management. Currently, there are no biomarkers to predict chemoresistance. Plasma gelsolin (pGSN) in circulating small extracellular vesicles (sEV) has previously been shown to predict chemoresistance in treatment-naïve EOC. Here, we expand upon sEV-pGSN as biomarker by incorporating MRI-based radiomics to improve the prediction of chemoresistance in EOC patients.

methodsIn this retrospective study, we used serum from 37 EOC patients with paired baseline MRI from the University of Hong Kong between 2016 and 2020. sEVs were isolated from serum samples using differential centrifugation and characterized by nanoparticle tracking analysis, western blotting, and transmission electron microscopy. Total pGSN and sEV-pGSN were quantified using sandwich ELISA. Radiomic features were extracted from the primary tumour on the MRI T2-weighted images (T2), apparent diffusion coefficient (ADC) maps (b = 0,400,800 s/mm2), and post-contrast images (PC). Highly correlated features (Spearman correlation coefficient of > 0.85) were removed and repeatable features selected using elastic-net regression. Grid-search 10-fold SCVs was utilized to optimize the hyper-parameters of the K-Nearest Neighbor (ADC and T2 + ADC + PC), Gaussian Naïve Bayes (T2), Linear Discriminant Analysis (PC), and Support Vector Machine (T2 + ADC) classifiers to build the prediction models, including total and sEV-pGSN.

resultsAmong the 37 EOC patients (56±11 years old), 65% presented at advanced stage (FIGO III-IV, n = 24). Thirty-one patients were chemosensitive and six were chemoresistant (progression free interval < 12 months). The combination of total and sEV-pGSN could predict chemoresistance (AUC = 0.591), however the inclusion of MRI radiomic features improved the test performance. The prediction model based on total pGSN, sEV-pGSN, and 4 selected T2 radiomic features showed the best performance in predicting chemoresponsiveness with the following mean performance metrics: AUC (0.973), sensitivity (0.833), specificity (0.968) and accuracy (0.946).

conclusionOur prediction model using total and sEV-pGSN and T2 features demonstrated excellent diagnostic ability in predicting chemoresistance in EOC patients, which could be used to facilitate alternate tailored therapeutics. Building on this work in larger multicentre studies will further validate these findings and clarify the utility of a combined radiomics/EV biomarker approach to chemoresistance prediction in EOC.