Patient-specific forecasting of prostate cancer progression during active surveillance using biomechanistic models and hybrid classifiers
Please login to view abstract download link
Active surveillance (AS) is a clinical management option for nearly 70% of newly-diagnosed prostate cancer (PCa) cases exhibiting low to intermediate clinical risk. AS delays definitive treatment until progression to higher risk is observed leveraging longitudinal multiparametric magnetic resonance imaging (mpMRI), serum prostate-specific antigen (PSA) tests, and biopsies. However, AS is based on an observational population-based approach, which largely ignores patient-specific tumor dynamics and can result in late detection of tumor progression. To address these issues, we propose using personalized predictions of PCa growth and progression informed by clinical and mpMRI data collected during AS. We use a biomechanistic model for tumor forecasting, which describes spatiotemporal PCa growth in terms of tumor cell density dynamics. The model is posed in the prostate geometry of the patient, which is segmented on T2W MRI data. We use apparent diffusion coefficient (ADC) maps obtained from diffusion-weighted MRI data to obtain estimates of tumor cell density. Model calibration and validation aim at assessing the model-data mismatch in spatiotemporal predictions of tumor cell density to find personalized model parameters and study the model predictive performance, respectively. Towards these ends, our patient-specific forecasts are obtained via model simulations that rely on isogeometric analysis (IGA). By using a logistic classifier, we further use model-based biomarkers calculated from the personalized forecasts (e.g., tumor volume, total proliferation activity) to predict PCa progression to higher-risk disease. Our pilot study in a small cohort of PCa cases (n=16) that were imaged three times in AS resulted in a concordance correlation coefficient (CCC) for tumor volume of 0.89 both at the time of the second and third mpMRI scans (i.e., calibration and prediction time horizons). The corresponding spatial fits and forecasts of tumor cell density maps produced a median CCC of 0.59 and 0.60, respectively. The logistic classifier of clinical risk achieved (i) an area under the receiver operating characteristic curve of 0.90, (ii) optimal sensitivity, specificity, and accuracy of 86.7%, 93.8%, and 90.3%, (iii) early PCa progression detection by more than 1 year with respect to standard AS. Thus, our PCa forecasting technology shows promise to build a digital twin to guide clinical decisions for each patient during AS.
