ESTRO 2024 - Abstract Book

S4578

Physics - Machine learning models and clinical applications

ESTRO 2024

1 South Florida Proton Therapy Institute, Medical Physics, Delray Beach, USA. 2 Medstar Georgetown University Hospital, Radiation Oncology, Washington D.C., USA. 3 South Florida Proton Therapy Institute, Radiation Oncology, Delray Beach, USA

Purpose/Objective:

This study assesses the effectiveness of incorporating surface mapping scores into machine learning (ML) algorithms for decision support in predicting current resimulation decisions in prostate cancer proton therapy. It analyzes 1767 sets of surface mapping (SM) scores from 2606 beams among 38 patients who underwent resimulation, as well as 4212 beams from 109 patients completing the standard course. Our novel SM tool was enhanced for efficiency and to provide SM scores for patients with non-spheroidal tumor volumes. The primary objective is to develop a novel quantitative method for standardizing resimulation decisions while also identifying the optimal algorithm for predicting these decisions using enhanced SM scores and patient metadata.

Material/Methods:

We enhanced our novel Surface Mapping (SM) tool for SM score calculation 1 . Based on SM score and affiliated patient and fraction metadata features, we trained multiple ML algorithms including Multilayer Perceptron Neural Network 2 , Random Forest 3 , Gradient Boosting 4 , and Decision Tree.

Metadata and patient CBCT images are obtained from the Varian Aria database. Patients with CBCT-guided IMPT for prostate cancer, undergoing repeated simulation CT scans with plan revisions, are the focus.

DICOM files of CBCT-Simulation CT pair voxel arrays are realigned to the same coordinate system via the registration CT. Treatment target contours obtained from the plan CT are projected from the simulation CT isocenter to the beam source via proton beam eye-view. After superimposing the planned target volume onto the CBCT coordinate array, a Delaunay and Poisson surface reconstruction algorithm renders the patient body surfaces of each CT scan. We improved on the previous algorithm by reorienting the coordinate frame around the axes of the plane defined by the beam-to-isocenter normal vector. Geometric deviations between two CT body surfaces within the projected proton field are calculated along each ray cast, employing a 0.25-mm analytical grid. Surface heatmaps for clinical evaluation are generated, as well as different percentile scores of that fraction’s CBCT from the Sim. For patients with multiple beams, the maximum scores at each percentile confer the final SM scores. The SM scores were combined with various metadata features extracted from the Aria database and used for training across multiple machine-learning algorithms. ML training utilized 70% of the data with grid search-optimized hyperparameters, while the remaining 30% was reserved for generating ROC curves and other processing. Subsequently, we derived feature importance using Gini importance metrics from these models.

ML model performances are compared against 98th and 95th percentile SM scores.

Results: