ESTRO 2024 - Abstract Book

S4573

Physics - Machine learning models and clinical applications

ESTRO 2024

There has been substantial interest in developing knowledge-based planning (KBP) pipelines to address the limitations presented by forward and inverse treatment planning across diverse radiation oncology applications 1,2 . Gamma Knife (GK) is a dedicated radiosurgery system when a large number of degrees of freedom that make it well suited to a KBP based planning approach. We have developed a novel GK-specific KBP pipeline utilizing 3-dimensional dose prediction in conjunction with inverse optimization (IO) for the generation of deliverable treatment plans. The efficacy of this pipeline was compared to plans resulting from both manual forward planning and a dose mimicking (DM) model. Furthermore, we evaluated the potential useability of the pipeline in clinical setting. Our results demonstrate that this pipeline consistently produces plans that achieve equal or superior quality within a similar time frame as state-of-the-art inverse planning methods.

Material/Methods:

MRI images, 3D dose distributions and target contour data was obtained for 349 patients treated for either brain metastases or intracranial schwannomas at Sunnybrook Health Sciences Centre. The data from 322 patients was modified using a GK-specific data modification method 3 , then used to train a neural network model for GK dose prediction. The trained model was then applied to predict dose distributions for 27 out-of-sample patients. Additionally, for these 27 patients, we acquired a general dose rate kernel, which was used to approximate dose rate kernels for each isocenter location. Subsequently, we developed a generalized IO model, based on an established inverse planning model 4 , to learn objective function weights from dose predictions. This model was solved using the obtained dose predictions and approximated dose kernels. The resulting weights were then used in the inverse planning model to generate deliverable treatment plans. The quality of the resulting KBP plans was compared to their clinical counterparts and plans resulting from a DM model using standard GK quality metrics and overall treatment time. Figure 1 provides an overview of the inputs and outputs of the optimization portions of both the IO and DM pipeline. Finally, we evaluated the overall average usage time of the pipeline and plan delivery characteristics to help determine its potential applicability in a clinical setting.

Figure 1: An overview of inputs and outputs in our a) inverse optimization and b) dose mimicking pipelines.

Results: