Machine learning-DeepSurv prediction model integrating mpMRI radiomics and genomic biomarkers for BCR-free survival and tumor response in prostate radiotherapy.
Authors
Affiliations (4)
Affiliations (4)
- Hezar Jarib Street, Isfahan University of Medical Sciences, Danesh Road, School of Medicine, Ground Floor, Student Research Committe & 1st Floor, Medical Physics Department, 2nd Room, Isfahan, Iran.
- Hezar Jarib Street, Isfahan University of Medical Sciences, Danesh Road, School of Medicine, 1st Floor, Medical Physics Department, Head Management Room, Isfahan, Iran.
- Hezar Jarib Street, Isfahan University of Medical Sciences, Danesh Road, School of Medicine, Alzahra Hospital, Radiology Department, Isfahan, Iran.
- Arak University of Medical Sciences, School of Medicine, 1st Floor, Anatomy Department, Arak, Iran.
Abstract
The purpose of this study was to design a radiogenomics machine learning-DeepSurv model for biochemical recurrence-free (BCR-free) survival and treatment response (TR) prediction for radiotherapy (RT) of prostate cancer (PCa). In this study, radiomic features were extracted from pre and post treatment multiparametric MRI (mpMRI), including T2-weighted (T2W), diffusion-weighted MR imaging (DWI) and apparent diffusion coefficient (ADC). Also, genomic biomarkers such as Ki-67 (a cell proliferation marker reflecting tumor growth activity and also prognostic information in cancer progression), PTEN (tumor suppressor gene regulating cell growth and survival, have a prominent role for TR and cancer progression) and Decipher (a genomic signature predicting cancer recurrence risk and TR based on gene expression patterns) were collected. Radiomics feature selection and dimensionality reduction methods were employed, followed by training machine learning (ML) models. Moreover, time to event data and survival models including Random Survival Forest (RSF) and DeepSurv neural networks were used. For model performance, the concordance index (C-index) and integrated Brier score (IBS), and for improving interpretability, the SHapley Additive exPlanations (SHAP) were applied. Radiomic feature of MRI including Kurtosis demonstrated a near-perfect positive correlation with Ki-67 expression (r = 0.64), however skewness showed a strong negative correlation with PTEN status (r = -0.88). Entropy and kurtosis of MRI were also highly correlated with the Decipher genomic risk score (r = 0.90 and r = -0.96, respectively). The integrated ML-DeepSurve model performance overall F1-score was 0.93 for TR. The model also offered robust stratification for patients based on BCR-free survival probability. Our findings underscore the potential of radiogenomic signatures as non-invasive biomarkers to personalized PCa RT decisions and provide a novel clinically explainable predictive model based on radiomic and molecular biomarkers for BCR-free survival and TR of mentioned cancer.