Background/Objectives: Extracting spatial features (texture analysis) from dose distributions (dosiomics) for outcome prediction is a rapidly evolving field in radiotherapy. To account for fraction size differences, the biological effective dose (BED) is often calculated. We evaluated the impact and added value of the BED in the dosiomics prediction modelling of grade ≥ 2 late rectal bleeding (LRB) probability within 5 years after treatment in three parts. Methods: For N = 656 prostate cancer patients previously treated in a randomized trial with conventional (CF) or hypofractionated (HF) radiotherapy, 42 dosiomic features were extracted from the dose distributions of the delineated rectum in physical doses and from dose distributions converted to the BED. Part 1: To assess whether an HF BED dosiomics model is generalizable to CF and vice versa, multivariate logistic regression BED models were constructed for HF and CF separately and tested on the other fractionation scheme. Part 2: The BED models were fitted to combined HF and CF data together to test whether this resulted in better models. Part 3: Separate physical HF and CF models were constructed and compared to the BED models. Results: Part 1: Dosiomics related to large-zone and long-run high-dose levels were predictive for both HF and CF. Deviation from the mean gray level was only predictive for HF. The BED HF model calibrations with CF data and vice versa were generally poor. AUCs ranged from 0.55 to 0.65. Part 2: Compared to the separate models, the models fitted to the combined HF and CF data showed better discriminative ability in CF but not in HF. Part 3: The apparent performances of models for the BED and physical dose were similar. Conclusions: Using the BED in the predictive dosiomic modelling of late rectal bleeding after prostate cancer radiotherapy to account for differences in fraction doses was of limited value.
Read full abstract