ESTRO 2024 - Abstract Book

S4956

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

145

Digital Poster

Radiomics with accumulated delivered dose to enhance rectal toxicity prediction in prostate cancer

Zhuolin Yang 1,2,3 , David J Noble 3,4 , Leila Shelley 1 , Thomas Berger 1 , Raj Jena 5 , Duncan B McLaren 3,4 , Neil G Burnet 6 , William H Nailon 1,2,3 1 Edinburgh Cancer Centre, Department of Oncology Physics, Edinburgh, United Kingdom. 2 University of Edinburgh, Institute for Imaging, Data and Communications, School of Engineering, Edinburgh, United Kingdom. 3 University of Edinburgh, Edinburgh Cancer Research, CRUK Scotland Centre, Institute of Genetics and Cancer, Edinburgh, United Kingdom. 4 Edinburgh Cancer Centre, Department of Clinical Oncology, Edinburgh, United Kingdom. 5 The University of Cambridge, Department of Oncology, Cambridge, United Kingdom. 6 The Christie NHS Foundation Trust, Proton Beam Therapy Centre, Manchester, United Kingdom

Purpose/Objective:

Although modern techniques such as IMRT have reduced severe toxicity events in prostate radiotherapy, predicting toxicities remains pivotal for patient-specific treatment. This study investigated the predictive power of combined radiomic features, from planning CT and daily MVCTs, along with planned dose and accumulated delivered dose information, for rectal toxicity prediction after prostate radiotherapy, including prediction of lower grade toxicities.

Material/Methods:

186 prostate cancer patients from the VoxTox study (UK-CRN-ID-13716) were investigated, all of whom underwent helical IGRT on TomoTherapy with 74 Gy in 37 fractions (N = 110) or 60 Gy in 20 fractions (N = 76), coupled with daily image guidance imaging. Toxicity data including grade ≥ 1 and grade ≥ 2 haemorrhage and proctitis (CTCAE v4.03), grade ≥ 1 stool frequency (LENT/SOMA) was prospectively recorded 2 years after radiotherapy. The 60 Gy data was converted to the equivalent 74 Gy dose in 37 fractions of 2 Gy. EUD and dose-width data, the maximum lateral extent of an ellipse fitted to the isodose cluster, were calculated from the planned and accumulated 2D-DSMs for each patient. On the pre-treatment planning CTs, IBSI-compliant radiomic features were extracted from subimages of size 8x8 pixel 2 on the rectal wall, which were averaged across all subimages and slices. For daily MVCTs, the treatment course for each patient was divided into three equal phases. MVCT radiomic features were derived within each phase, employing the same methodology used for the planning CT radiomic feature extraction. Patients were split into a training (75%, N = 139) and a test (25%, N = 47) set. Mann-Whitney U test followed by a Spearman’s rank correlation were used to identify potential predictors and remove highly correlated features. These predictors were fed into a multivariate logistic regression model with elastic net penalty. Hyperparameters were fine tuned using 5-fold cross validation. AUC values were reported with 95% CIs from 100 bootstraps. Predictive performance of models with different feature combinations was compared by one-sided Mann-Whitney U rank tests based on AUC values. Fig.1 shows the main workflow of the analysis.