Room: Exhibit Hall | Forum 6
Purpose: To identify and optimize the choice of region of interest in optical surface monitoring for deeper tumor using a machine learning based framework, and to evaluate its feasibility in clinical application.
Methods: We retrospectively picked a patient who received 30 fractions of pelvis radiation treatment. The patient was simulated prune on CT sim. The pre-treatment setup included an initial positioning based on triangulation followed by daily whole pelvis cone beam computerized tomography (CBCT). Setup corrections were made based on online registration of the GTV between CBCT and the planning CT. The vertical shift with regard to the triangulation were recorded for each fraction. A regression module was created using Tensorflow, taking the parameterized patient posterior body surface as the training set (leave one out for testing), and the vertical shifts as the output to predict. To determine which part of the body surface is optimal to predict the vertical shift, the body surface was divided into multiple subsets and feed into the regression module separately.
Results: The daily vertical shifts ranged from -1.4cm to 1.1cm relative to the triangulation with an average shift of -0.2cm. Regression using the whole body surface yield an average estimation error of 0.4cm. One of the subset of the body surface, which consisted surface of inner thighs, predicted the vertical shift with an average error of 0.23cm, was the best predictor among all subsets of the body surface.
Conclusion: For the patient in the study, a machine learning based regression module established a correlation between body surface and deeper tumor position with acceptable accuracy. The framework also helped to identify the optimal region on the body surface to predict the tumor position. It can be of clinical relevance for some patients when use together with optical surface monitoring system to track deep tumor motion.