Room: AAPM ePoster Library
One of the more difficult problems associated with building large DICOM-RT databases is properly identifying regions-of-interest (ROIs) that were not consistently named. Prospectively, this problem can be addressed with standardized nomenclature, however, large-scale retrospective studies could greatly benefit from a more automated and robust method.
As a proof-of-concept, we demonstrate feasibility of transforming ROI surface coordinates into a tensor, suitable for machine learning, to identify anatomical structures based on DICOM-RT structure coordinates. These tensors are calculated using the ROI’s centroid (self-centroid), a second ROI’s centroid (paired-centroid), or paired surface points of a second ROI (paired-surface) as origins. An algorithm was developed taking advantage of both the self-centroid method and a paired method where several passes of self-centroid results inform the paired method. Using a random forest classifying algorithm, models were generated and tested using a Head & Neck data set from a single physician, including 70 laryngeal patients and 1,699 anatomical ROIs among 30 ROI categories.
Overall accuracies and confusion matrices were generated for each tensor method, using the larynx as the paired ROI for the paired methods. Overall accuracies of 88.5%, 95.3%, and 93.8% were reported for the self-centroid, paired-centroid, and paired-surface methods, respectively. However, an overall accuracy of 100% with a single rejected ROI was observed when combining the self-centroid and paired-centroid methods – without explicitly identifying the “paired” ROI.
Feasibility of automated ROI identification was demonstrated without the need for complicated and computationally expensive neural networks. This was accomplished by transforming DICOM-RT Structure coordinates with a map projection for input into a random forest machine learning algorithm.
DICOM-RT, Deformation, Structure Analysis
IM/TH- Mathematical/Statistical Foundational Skills: Machine Learning