S2476
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
approximately 20-minutes apart, mimicking a fraction. Clinical target volume (CTV) and gross tumour volume (GTV) were segmented on each image by two expert observers. Images were pre-processed using histogram normalisation. Radiomic features (n=107) were extracted using PyRadiomics (v3.0.1) across all segmentations and images for CTV and GTV. Repeatability was assessed by comparing the intra- fraction images using the intra-class correlation coefficient ICC [1,2], with contours as raters and scans as targets [3]. Sensitivity was assessed by the relative percentage change in feature value across visits from baseline. A confidence ratio was calculated as the inter-fractional percentage change, reflecting MDC, divided by the intra-fractional percentage change, reflecting repeatability, to show the MDC to repeatability relationship.
Results
75% (CTV) 62%(GTV) of features showed excellent (ICC > 0.9) repeatability and 1% (CTV), 3% (GTV) features were labelled as poorly repeatable (ICC < 0.5), Figure 1. There was a systematic increase of 5% in the CTV feature repeatability over time (Figure 1a), but not in the GTV features (Figure 1b). The ratio between MDC (inter-fractional change) and repeatability (intra- fractional change) across time is shown in Figure 2. Over 57% of features showed a greater inter-fractional percentage change than intra-fractional percentage change suggesting these features are suitable for supervised feature selection with tumour response.
Conclusion Feature repeatability is not consistent throughout the course of radiotherapy. Features derived from the CTV exhibit greater stability than those derived from the GTV, likely due to the larger calculation volumes involved. Selecting features based solely on repeatability at baseline may inadvertently exclude
Made with FlippingBook - Share PDF online