Can Dosiomics Features Be Relevant Predictive Factors for Biochemical Recurrence After Radiotherapy in Prostate Cancer Patients

PURPOSE/OBJECTIVE(S) Radiotherapy is one of the main treatment strategies for localized prostate cancer. Although several patients receive external radiotherapy, the relationships between dosimetric factors related to the treatment planning and biochemical recurrence (BCR) remain unclear. The aims of the present study were to develop a prediction model for BCR by integrating dose-based radiomics (dosiomics) features and clinical variables in prostate cancer patients, and to clarify the relationships between the dosimetric factors and BCR. MATERIALS/METHODS A total of 489 patients with adenocarcinoma of the prostate who had been prescribed a dose of 78 Gy in 39 fractions with intensity modulated radiation therapy from 2007 to 2014 were retrospectively enrolled. A total of 2,475 dosiomic features were calculated from the 3D doses in prostate, clinical target volume (CTV), and planning target volume (PTV), and relevant 8 clinical variables for BCR were also considered. Feature selection was performed to eliminate redundant or not significant features. Multivariate cox proportional hazards regression model was trained on the training cohort of 342 patients using the selected features, and the predictive performance was validated on the validation cohort of 147 patients using the concordance index (C-index). Kaplan-Meir curves with log-rank analysis was used to assess the univariate discrimination of freedom from biochemical failure (FFBF) between high- and low-risk BCR groups. RESULTS A total of 96 patients developed BCR with the median time of 55.2 months. Two dosiomic features (1 from CTV and 1 from PTV) and three clinical variables were selected as predictive factors. Of the five, CTV (HGLRE) (hazard ratio [HR]: 0.73; 95% confidence intervals [CI]: 0.57-0.93; P = 0.01), pretreatment prostate-specific antigen (HR: 1.27; 95% CI: 1.04-1.55; P = 0.02) and positive biopsy core rate (HR: 1.40; 95% CI: 1.07-1.82; P = 0.01), were significantly associated with BCR. The C-index of the model in the validation cohort was 0.67 (95% CI: 0.65-0.68). CTV_wavelet-HHH_glrlm_HGLRE could significantly distinguish the patients into high- and low-risk groups (8-year FFBF: 76.6% vs. 87.5%, P < 0.01). CONCLUSION The dosiomic feature extracted from CTV can be a relevant predictive factor for BCR after radiotherapy in prostate cancer patients and may be useful as new metric for evaluating the quality of a treatment plan, instead of conventional dose indices.