Cone Beam CT Based Delta-Radiomics of Prostate Cancer Patients and Their Correlation to Quality of Life, Genitourinary and Gastrointestinal Toxicities

Abstract

For patients receiving external beam radiotherapy (EBRT) for prostate cancer (PCa), acute and late radiation-related toxicities can significantly impact short and long-term quality-of-life (QOL). Radiomics, an analytic technique that extracts patterns from medical imaging, has been incorporated into predictive models in diverse cancers. Daily cone beam CT (CBCT) images during EBRT allow for the study of changes in radiomics over time, known as delta-radiomics. We hypothesize that delta-radiomics of the prostate extracted from daily CBCTs could be predictive of genitourinary toxicity (GUT), gastrointestinal toxicity (GIT), and patient QOL.From an IRB approved database, 31 PCa patients were identified who received EBRT with daily CBCTs. Patients were treated on dose fractionation schemes of 76-91 Gy in 37-40 fractions and some received a 12 Gy boost. The prostate was contoured on all daily CBCTs and 42 radiomics features were extracted for analysis. Radiomic features included intensity-based, gray-level run length (GLRLM), gray-level co-occurrence (GLCM), neighborhood gray-level difference (NGTDM), and gray-level size zone (GLSZM) matrix classes. These radiomic features were averaged per week to decrease data granularity. Patient outcomes analyzed included the total International Prostate Symptom QOL score GUT, GIT, combined GU/GIT grades. Logistic and linear regression models were used to estimate effects of radiomics features on later occurring toxicity and QOL, respectively. Odds ratios (ORs), regression coefficients, 95% confidence intervals, and P-value were estimated. Area under the receiver operating characteristic curve (AUC) and R2 were obtained as a measure of models' goodness-of-fit. The statistical significance threshold was P < 1.5e-04 using an alpha level of 0.05 and applying the Bonferroni correction for multiple comparisons. All tests were two-sided. Parallel data sets were generated using standard (sCBCT) and iterative (iCBCT) CBCT reconstructions. Radiomic feature were preprocessed using Lloyd quantization with Collewet normalization.Five CBCT delta-radiomic features significantly correlated on week two with QOL occurring after. The results are summarized in Table 1. The R2 values indicate a moderate effect size. GUT and GIT were not found to correlate with the radiomic features. All significant radiomic features corresponded to sCBCT.Future short-term changes to QOL were detected as early as week two with CBCT delta-radiomics. However, the GUT, GIT, and GU/GIT toxicities were not predicted. These results can serve as preselection criteria of radiomic features to develop more sophisticated machine learning models with larger sample and improve statistical power.