Decision Forests with Spatio-temporal Features for Graph-based Tumour Segmentation in 4D Lung CT

We propose an automatic lung tumor segmentation in dynamic CT images that incorporates the novel use of tumor tissue deformations. In contrast to elastography imaging techniques for measuring tumor tissue properties, which require mechanical compression and thereby interrupt normal breathing, we completely avoid the use of any external physical forces. Instead, we calculate the tissue deformations during normal respiration using deformable registration. We investigate machine learning methods in order to discover the spatio-temporal dynamics that would help distinguish tumor from normal tissue deformation patterns and integrate this information into the segmentation process. Our method adapts an ensemble of decision trees combined with a 3D graph-based optimization that takes into account spatio-temporal consistency. The experimental results on patients with large tumors achieved an average F-measure accuracy of 0.79.