Purpose: Side effects of prostate cancer (PCa) treatment have the potential to impact one in eight US men. During intensity modulated radiotherapy (IMRT) for PCa, cone-beam CT (CBCT) images are acquired daily to improve the accuracy of treatment delivery. Delta-radiomics applied to PCa CBCT images can aid in prediction of clinical endpoints such as genitourinary toxicities (GU-tox) and changes in International Prostate Symptom Scores (delta-IPSS). Since the predictive power of delta-radiomics increases with sample size, the need for manual delineation of anatomic structures on every PCa CBCT is a limiting factor in model development. The purpose of this study is to assess differences in predictive performance of delta-radiomic features extracted from CBCT images with manual vs automatically segmented prostate volumes.
Methods: Twenty-one patients receiving IMRT for PCa were included. CBCT images were reconstructed based on parameters previously found to improve repeatability and reproducibility of radiomic features. Prostate volumes were manually contoured and automatically generated for all fractions of all patients using deformable image registration. Forty-two radiomic features plus eight volume-normalized alternatives were extracted from contoured volumes. Radiomic data were averaged into 20 Gy BED bins and normalized to the first fraction. Delta-radiomic logistic regression models for GU-tox and delta-IPSS correlation were created using features preselected with random forest. The Delong test was used to evaluate significance between the two model sets with a threshold p<0.05.
Results: Delta-radiomic models using radiomic features derived from the automatically generated and manually contoured prostate volumes were not found to be significantly different at all timepoints for either GU-tox or delta-IPSS.
Conclusion: Delta-radiomic model performance is not significantly different between manually or automatically contoured prostate volumes on daily CBCT images. Thus, automatically generating the prostate is a suitable replacement for manual contouring for CBCT-based delta-radiomics. This will aid future studies in accumulating larger sample sizes.
Funding Support, Disclosures, and Conflict of Interest: Acknowledgments This work was supported in part by a research grant from Varian Medical Systems, Inc., Palo Alto, CA. (GR013242). Conflict of Interest Dr. Abramowitz has received an honorarium from Varian Medical Systems. The other authors have no relevant conflicts of interest to disclose.
Cone-beam CT, Quantitative Imaging, Deformation