ESTRO 2024 - Abstract Book

S4541

Physics - Machine learning models and clinical applications

ESTRO 2024

Results:

A.1. Shape descriptor variations: A comparison of the resolution-based shape descriptors using various cut-off values and their respective reconstruction accuracy measured by the Dice coefficient showed smaller grid dimensions and a higher number of target points might improve the quality of the descriptors and their reconstructions. This yields an average accuracy of 95% for all involved organs, with better (and more coherent) results for the bladder and the prostate. For more accurate and reliable organ representations, we propose a fine grid resolution (e.g., 10 mm grid resolution). A.2. Data missingness/incompleteness: Our results showed that this method is robust for a low number of missing layers, where if 2-3 slices are missing the reconstruction accuracy still reaches 95% for all organs. However, missing slices in the core of the organs have an impact on the outcomes of the upsampling and reconstruction workflow. B.1. Different clustering methods: AHC provides a way to cluster patients, but other clustering methods can be alternatively used. Among all the alternatives we investigated, the clustering results are overall consistent. B.2. Clustering parameterizations: Our analysis has highlighted that different settings (distance metric and linkage) yield only slightly different clusters. Regarding the distance metric, desired ranking patterns are marginally best achieved by the Euclidean distance, while linkage does not have a big impact with the exception of average linkage, which tends to produce irregular results. B.3. Sensitivity: Including additional patient data within the clustering step does not influence the outcomes, indicating that the stability of cluster assignments is not influenced by possible outliers.

Conclusion:

Our study assessed models for predicting pelvic organ variability in radiotherapy. We placed emphasis on the impact of organ shape descriptors and clustering on predictive outcomes. Our findings highlight that the pivotal factor is the choice of shape description for robust anatomical variability predictions. Our future research aims to link further anatomical variability prediction with treatment outcomes, and gastrointestinal and genitourinary toxicity attributable to radiotherapy.

Keywords: anatomical variability, prediction robustness

References:

[1] Furmanová, K., Muren, L.P., Casares-Magaz, O., Moiseenko, V., Einck, J.P., Pilskog, S. and Raidou, R.G., 2021. PREVIS: Predictive Visual Analytics of Anatomical Variability for Radiotherapy Decision Support. Computers & Graphics, 97, pp.126-138.

2516

Mini-Oral

NTCP model of acute toxicity after whole breast irradiation: training and extended validation