ESTRO 2024 - Abstract Book

S4540

Physics - Machine learning models and clinical applications

ESTRO 2024

2507

Digital Poster

Assessing the impact of shape descriptors and clustering on predicting RT pelvic organ variability

Ádám Böröndy 1 , Katarína Furmanová 2 , Oscar Casares-Magaz 3 , Vitali Moiseenko 4 , John P. Einck 4 , Renata Georgia Raidou 1 1 TU Wien, Research Unit of Computer Graphics, Vienna, Austria. 2 Masaryk University Brno, Department of Visual Computing, Brno, Czech Republic. 3 Danish Center for Particle Therapy, Department of Clinical Medicine, Aarhus, Denmark. 4 UC San Diego, Department of Radiation Medicine and Applied Sciences, La Jolla, USA

Purpose/Objective:

Our recent research delved into the analysis and prediction of variations in pelvic organ anatomy, as these anatomical features and their variations relate to gastrointestinal and genitourinary toxicity, attributable to radiotherapy. In particular, PREVIS [1] leveraged a retrospective patient cohort with established variability to tailor predictions for new patients. PREVIS involves four key steps: (1) Conversion of organ shapes to mathematical descriptors, (2) Summarization of shape variability for individual patients as a deviation from the mean shape computed from all CT scans, (3) Stratification of the cohort by means of clustering, and (4) Predictions about new patients using precedents provided by clustering. The accuracy of the predictions made with PREVIS hinges on various statistical and machine learning techniques that require rigorous evaluation. The purpose of the present study is to validate the robustness of anatomical variability prediction in radiotherapy and to assess uncertainties caused by alternative choices within the outlined workflow. We conducted a leave-one-out cross-validation within a retrospective matched-case control cohort of 29 prostate cancer patients who received conventionally fractionated radiotherapy with a daily CBCT-based patient setup. Planning CT, five consecutive CBCTs from the first week of treatment, and one CBCT from each of the remaining treatment weeks were included in the analysis. The data were used to rigorously explore the following aspects: A. Shape descriptors: Methods were developed to compare bladder, rectum, and prostate shape descriptors focusing on reversibility, scalability to new patients, size control, and shape comparability. A.1. Shape descriptor variations: Vectors were constructed by considering attributes such as bounding box calculations and probability assignments to key points within the organ space. This enables flexible shape reconstruction at different resolutions. The impact of organ centering relative to the prostate and the aggregation of descriptors was assessed over multiple time steps. A.2. Data missingness/incompleteness: The impact of omitting slices in organ delineation on the accuracy of the shape descriptor was assessed. B. Clustering: Clustering techniques were used to stratify the cohort according to patient similarity. B.1. Different clustering methods: Beyond the agglomerative hierarchical clustering (AHC) utilized by PREVIS, alternatives, including k-means, k-medoids, model-based clustering, and fuzzy c-means were explored. B.2. Clustering parameterizations: Various distance measures and linkage criteria for AHC were investigated. This includes Euclidean distance and complete linkage, as well as Manhattan, Minkowski, Canberra, Binary, and Maximum distance, combined with single, average, McQuitty, median, centroid, and Ward linkage. We also analyze how the number of clusters impacts our findings. B.3. Sensitivity: Cluster sensitivity to the late inclusion of additional patient data was probed. Material/Methods: