Automatic Semantic Parsing of CT Scans via Multiple Randomized Decision Trees

PURPOSE We introduce a new, efficient algorithm for the automatic detection and localization of anatomical structures within 3D CT images. Our algorithm builds upon recent randomized decision tree classifiers and produces accurate posterior probabilities for each of the classes (e.g. organ labels) in the training set. Accurate results are obtained by exploiting the high level of generalization offered by the classifier. Furthermore, its massive parallelism yields high computational efficiency. Multiple decision trees (different from one another) are trained on labelled data to detect and localize anatomical structures within 2D or 3D scans; independently of their resolution, cropping and other common transformations. During testing the trees are applied to previously unseen images and the relevant anatomical structures automatically recognized and segmented. 3D spatial context is modelled by means of efficient visual features built upon integral volume processing. The output of our algorithm is probabilistic thus allowing uncertainty modelling as well as fusion of multiple sources of information. Our algorithm does not necessitate the use of atlases, with all the issues that that entails. RESULTS A database of labelled CT images has been split into training and testing. Quantitative results have been reported on the testing subset only, to make sure the algorithm learns to generalize. Excellent accuracy (localization accuracy of around 2 cm) has been observed in detecting organs such as liver, heart, kidneys, lungs, eyes, and head. Comparison with state of the art algorithms such as SVM and GMMs show superior performance for our technique.