Radiation hematologic toxicity prediction in rectal cancer: a comparative radiomics-based study on CT image and dose map.

Abstract

Background and objectives: Acute radiation hematologic toxicity may disturb the radiotherapy plan and thus decrease the treatment outcome. However, whether the dose map has enough prediction value for detecting hematologic toxicity (HT) is still unknown.

Methods: In this study, the pre-treatment CT images and the in-treatment dose map were collected from a discovery dataset of 299 patients and a validation dataset of 65 patients from another center. Then, the radiomic features of the clinical target volume (CTV) in the radiotherapy were extracted, and the least absolute shrinkage and selection operator (LASSO) algorithm was used for feature dimension deduction; three classifiers, that is, support vector machine (SVM) (rbf kernel), random forest, and CatBoost, were used to construct the HT classification model in rectal cancer patients. The model performance was evaluated by both the internal 20% dataset and the external multicenter dataset.

Results: The results revealed that CatBoost achieved the best model performance in almost all tasks and that CT images performed similarly with the dose map, although their combination model performed lower. In addition, gender, age, and some radiomic features from the decomposed image space were the most representative features for HT prediction.

Conclusion: Our study can confirm that the HT occurrence in locally advanced rectal cancer (LARC) patients was multifactorial, and combining effective features together can classify the high-risk patients with HT, thus timely preventing or detecting HT to improve the subsequent outcome.