Towards unification of organ labeling in radiation therapy using a machine learning approach based on 3D geometries

Abstract: In radiation therapy, it is important to control the radiation dose absorbed by Organs at Risk (OARs). The OARs are represented as 3D volumes delineated by medical experts, typically using computed tomography images of the patient. The OARs are identified using user-provided text labels, which, due to a lack of enforcement of existing naming standards, are subject to a great level of heterogeneity. This condition negatively impacts the development of procedures that require vast amounts of standardized data, like organ segmentation algorithms and inter-institutional clinical studies. Previous work showed that supervised learning using deep-learning classifiers could be used to predict OARs labels. The input of this model was composed of 2D contours of the OARs, while the output was a standardized label. In this work, we expanded this approach by qualitatively comparing the performance of different machine learning algorithms trained on a clinical data set of anonymized prostate cancer patients from the Iridium Kankernetwerk clinic (Belgium). The data set was partitioned in a semi-automatic fashion using a divide-and-conquer-like approach and various 2D and 3D encodings of the OARs geometries were tested. Moreover, we implemented a reject class mechanism to assess if the inference probability yielded by the model could be used as a measure of confidence. The underlining goal was to restrict human intervention to rejected cases while allowing for a reliable and automatic standardization of the remaining ones. Our results show that a random forest model trained on simple 3D-based manually engineered features can achieve the twofold goal of high classification performance and reliable inferences. In contrast, 3D convolutional neural networks, while achieving similar classification results, produced wrong, but confident, predictions that could not be effectively rejected. We conclude that the random forest approach represents a promising solution for automatic OAR labels unification, and future works should investigate its applications on more diversified data sets.