S2492
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
up. This study explores the feasibility of semi- automatically segmenting the vaginal wall and extracting the 3D dose distribution in that region, as well as using dosiomic, radiomic, and clinical variables to predict vaginal stenosis. Material/Methods: Clinical variables, 3D dose grids, and planning MRIs were retrospectively collected for 93 cervical cancer patients treated with pulsed-dose-rate BT between 2015 and 2023. Vaginal stenosis was scored using CTCAE V4.0, grouping patients by grade <2 (n=59) and ≥ 2 (n=34). The vaginal wall was semi-automatically segmented after percentile-based intensity rescaling, using transverse MRI slices from ring centre to PIBS point. Segmentations were resampled to the 3D dose grid for spatial alignment and dose values within this region were converted to equivalent dose in 2Gy fractions (EQD2 α / β 3). As a feasibility assessment, 3D dose-volume parameters (D0.1cm ³ , D1cm ³ , D2cm ³ , D50%) of the total dose in EQD2 α / β 3 (EBRT + BT) to the vaginal wall were compared between the two groups. Dosiomic and radiomic features were extracted from original, wavelet-transformed and Laplacian-of-Gaussian filtered segmentations using PyRadiomics. Highly correlated features were removed prior to modelling. Within each fold of 20 repetitions of 5-fold stratified cross-validation (CV), logistic regression models with embedded LASSO feature selection were constructed separately on dosiomic, radiomic, and clinical features. Selected features from each model were then combined within the same fold in an additional logistic regression model. Model performances were evaluated using AUC. The full pipeline is shown in Figure 1.
Results: On average, all 3D dose-volume parameters were higher in patients with grade ≥ 2 vaginal stenosis. However, substantial variability was observed in both groups, with only D50% showing a statistically significant difference (see Table 1). After correlation filtering, 192 of 1116 dosiomic, 343 of 1130 radiomic, and all 21 clinical features remained. Across repeated CV, test AUCs (average±standard deviation) were 0.64±0.12 (dosiomics), 0.71±0.12 (radiomics), 0.51±0.08 (clinical), and 0.65±0.13 (combined model). While model performance was modest, train AUCs were notably higher (up to 0.89±0.04 (combined model)). This indicates overfitting and the need for a larger sample size and more sophisticated machine learning approaches.
Conclusion: Semi-automatic segmentation and volumetric dose distribution extraction of the vaginal wall is feasible. Using associated dosiomic and radiomic features in a LASSO logistic regression model provided modest predictive power for grade ≥ 2 vaginal stenosis. Keywords: brachytherapy, dosiomics, vaginal stenosis
Made with FlippingBook - Share PDF online