Room: Track 3
The post-implantation dosimetry of the CivaSheet LDR device is a time-consuming task that requires manual reconstruction of both the locations and orientations of tens of individual directional sources (CivaDots) from a CT scan of the patient. We present an automated workflow that takes advantage of the regularities in the spatial distribution of the individual sources and results in improved speed and accuracy.
Since the exact number of implanted sources is known, a modified k-means clustering algorithm can be adopted for identifying the voxels comprising each CivaDot. The two-dimensional grid-like arrangement of the CivaDots is exploited in a Principal Component Analysis framework to identify the global orientation of the CivaSheet. Smooth bi-variate spline interpolation is subsequently applied to obtain the analytical shape of the entire sheet. The resulting 3D surface is guaranteed to be differentiable, allowing for accurate determination of both the normal vector at the location of each CivaDot and the local curvature of the CivaSheet. A modified TG-43 formalism is then used for dose calculation on the post-implantation CT scans of five patients.
The workflow was tested on the CT scans of two CivaSheets attached to surfaces of known shape (flat and cylindrical with 3.5cm diameter). The reconstructed normals deviated from the expected directions on average by 2.5 degrees and 1.2 degrees for the flat and cylindrical CivaSheets, respectively. For 5 treated patients the dose distribution from the automated algorithm agreed well (distance-to-agreement < 0.5mm for the prescription isodose line) with the manually reconstructed dose, calculated by an experienced user in BrachyVision TPS.
We present an automated workflow for post-implantation dosimetry of the CivaSheet LDR device. The implementation of unsupervised machine learning algorithms for identifying both the locations and orientations of the directional sources ensures improved speed and reproducibility compared to the manual workflow.
TH- Brachytherapy: Treatment planning using machine learning/automation