Integrated Spatio-Temporal Segmentation of Longitudinal Brain Tumor Imaging Studies

摘要

Consistent longitudinal segmentation of brain tumor images is a critical issue in treatment monitoring and in clinical trials. Fully automatic segmentation methods are a good candidate for reliably detecting changes of tumor volume over time. We propose an integrated 4D spatio-temporal brain tumor segmentation method, which combines supervised classification with conditional random field regularization in an energy minimization scheme. Promising results and improvements over classic 3D methods for monitoring the temporal volumetric evolution of necrotic, active and edema tumor compartments are demonstrated on a longitudinal dataset of glioma patient images from a multi-center clinical trial. Thanks to its speed and simplicity the approach is a good candidate for standard clinical use.