S2489
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Dataset 1 who underwent repeat scans were used for test-retest analysis.Limbus autocontour software (v1.8.1, Radformation, USA) was used to outline prostates on all scans. Pyradiomics (v3.1.0) [3] was used to extract radiomics features. Figure 1 demonstrates feature reduction. Each dataset was split 70:30 into training and test sets. The test set was further split 50:50 into unseen validation and temporary-testing subsets. An iterative training process was performed with sample swapping between the training set and temporary-test subset. Seven machine learning algorithms were used to build seven models each iteration. Training iterations were repeated until at least one model reached AUC >0.7. An ensemble VotingClassifier model was then used to combine each algorithm’s results to produce a baseline final model. Final models were tested on the 15% unseen validation set and the full external validation dataset.
Conclusion: Radiomics-based models can perform well on unseen local data, but are poor when validated externally. Imbalanced BCR rates, as well as different features being selected across datasets are likely to influence this. Use of autocontouring software eliminated differences in contouring practice as a confounder across datasets. References: 1. Osman et al. 2019 Int J Rad Oncol Biol Phys 105 (2): 448–562. Roach M et al. 2006. Int J Radiat Oncol Biol Phys. 65(4):965–74.3. Griethuysen et al. 2017. Cancer Research 77(21), e104–e107. Keywords: Prostate, recurrence, validation Contrast enhanced CBCT capabilities of Halcyon with HyperSight: moving towards functional CBCT Riley McMaster 1,2 , Raanan Marants 2 , Hedi Mohseni 2 , Catherine Coolens 1,2 1 Medical Biophysics, The University of Toronto, Toronto, Canada. 2 Medical Physics, Princess Margaret Cancer Centre, Toronto, Canada Purpose/Objective: Dynamic contrast enhanced CT (DCE-CT) is a functional imaging method for measuring vascular perfusion with high resolution and clinical value. CBCT has previously been limited by low image quality and lengthy acquisition times preventing valid analysis of Digital Poster Highlight 4621
Results: BCR rates differed between centres (Dataset 1: 22% BCR, median follow-up: 8.9 years; Dataset 2: 30% BCR, median follow-up: 5.8 years). Final model 1 returned AUC=0.75 (accuracy=0.79) for internal validation and AUC=0.50 (accuracy=0.70) when externally validated. Final model 2 had internal validation AUC=0.60 (accuracy=0.69), and external validation AUC=0.53 (accuracy=0.58). Final model 1 outperformed Final model 2 on their respective local validation datasets (Fig. 2A/B). This may be influenced by the test-retest cohort originating from the same centre as Dataset 1. There were 118 features selected for Final model 1 and 138 for Final model 2, with 98 common to both (Figure 2C).
Made with FlippingBook - Share PDF online