Radiomics and Mechanistic Hybrid Model for Prediction of Recurrence in Prostate Cancer

In prostate cancer radiotherapy, biochemical recurrence has been traditionally predicted using radiomics approaches with limited performance. The purpose of this work was to use mechanistic in silico simulation outputs and radiomics features, combined with clinical information, to improve the prediction of biochemical recurrence and to compare different model and feature selection strategies. A cohort of 213 patients with localized prostate adenocarcinoma having undergone external beam radiotherapy was used. Analogous digital tissues were built from radiomics features extracted from pre-treatment T2-w and ADC MRIs. The prescribed irradiation protocols were simulated using the mechanistic model. A grid search method was used to find the best combination of parameters to predict recurrence using : i) only clinical data, ii) only radiomics data from T2-w and ADC images, iii) only simulation outputs, iv) all data types together and v) a voting ensemble method. Significant improvement in prediction (p-value ≤ 0.01) was achieved using the combination of data collection (AUC=0.75) compared to prediction based on clinical data only (AUC=0.64), MRI features only (AUC=0.69), and the voting ensemble (AUC=0.71). No significant difference were observed compared to the mechanistic model (AUC=0.72).