PHSPP02 Presentation Time: 2:40 PM: Automated Channel Identification for Efficient Brachytherapy Treatment Planning

Abstract

<h3>Purpose</h3> For interstitial brachytherapy, channel reconstruction during the process of brachytherapy treatment planning can be an arduous and time-consuming task, which becomes more difficult and error-prone when 20 to 30 different catheters with curvatures are involved and overlap in 3D in the patient anatomy. The goal of this project is to develop an artificial intelligence-based approach for automated channel identification using patient CT scans, reducing both the time and potential for human error in brachytherapy treatment planning. <h3>Materials and Methods</h3> Three sets of CT scans were used as well as a baseline model for training, which were acquired with a set of binary-coded "dummy wires" markers <i>in situ</i>. Each of the 12 available dummy wires was designed with a unique pattern of radio-opaque markers and/or spacing strands which was mapped to a channel identification array. An identification algorithm was developed which involved first manually specifying the 3D DICOM coordinates of the centroids of each of the marker strands within the dummy wires. A nearest neighbor search was performed using a k-d tree method for each centroid acting as a potential starting point within a channel. Using the knowledge from the baseline model of each of the 12 potential dummy wire markers, the system predicted the most probable trajectory through the centroids. Then, to assign the proper channel number, the identified pattern of the applicable neighboring points was correlated with the channel identification array defined previously. This process was repeated until the entire configuration of channel identities was determined. Finally, a process of elimination was used to output the correct channel labels and locations. <h3>Results</h3> In this feasibility study, the developed algorithm was able to suitably provide channel identities and return the locations to the user when the proper centroid locations were input into the model. The model ran within seconds and required no training data other than the layout and spacing of the strands within the coded dummy-wire marker system. The final output was able to be graphically represented as labeled reconstructions of the catheters in a 3D space. <h3>Conclusions</h3> The outcomes of this preliminary work suggested that there is a future for automated aid in the channel identification process based on coded dummy-wire markers for more efficient treatment planning in brachytherapy. Using only the centroids of the markers, the channels could be accurately reconstructed. Future work will involve creating a 3D CT scan-based image detection algorithm to automate the process of defining the centroid locations without any <i>a priori</i> knowledge.