Abstract
Purpose/Objective(s)- Vaginal canal stenosis resulting from radiation-induced fibrosis affects up to 60% of locally advanced cervical cancer (LACC) patients and is often a major cause of long-term distress in cancer survivors. Known risk factors include tumor extension, RT dose, and vaginal treatment volume, while these factors are well-known, their predictive power remains insufficient without incorporating new predictive tools such as the use of radiomics. Radiomic analyses on the relationship between radiation dose distribution and toxicity have previously been performed. To our knowledge, however, analyses on possible relationships between radiomic features present in pre-treatment imaging and VS rates have never been performed.
Materials/Methods- 24 women with LACC treated with EBRT and subsequent brachytherapy between March 2016 and January 2023 were identified through a retrospective review of an institutional cervical cancer database. Patients were divided into two groups based on post-RT presence or absence of VS, as documented by clinical exam. Patients without detailed pelvic exam documentation were excluded. The vaginal canal was selected as the region of interest (ROI) and was delineated manually on pre-and post-treatment CT imaging and used for feature extraction using the PyRadiomics open-source Python package. For the purpose of this study, extraction was only performed on the pre-treatment CT simulation scan. The relationship between VS and each radiomic feature was evaluated using a Mann-Whitney U test, and radiomic features found to be significantly different between patients with and patients without VS were selected. To identify additional the most predictive radiomic features while minimizing overfitting, feature selection was then performed using a stepwise logistic regression between the presence of vaginal stenosis and the selected features.
Results- 13/24 patients had VS and 11/24 had no VS on post-RT physical exam. 107 radiomics features, including intensity (n = 18), shape (n = 14), and textural features (n = 65), were extracted from the ROI delineated on individual CT scans. On Mann-Whitney U testing, 34/107 features exhibited a statistically significant difference between the two groups (p values ranged from 0.001 to 0.042). Logistic regression identified 25 out of 34 features as significantly correlated with VS (p-values ranged from 0.001 to 0.246).
Conclusion- Radiomic features identified in this retrospective study can be used in predictive modeling of VS risk using pre- and post-RT images, which can be validated in future prospective studies. These would enable early identification of high-risk individuals for potential intervention strategies.