Resection-Based Demons Regularization for Breast Tumor Bed Propagation

A tumor resection introduces a problem of missing data into the image registration process. The state-of-the-art methods fail while attempting to recover the real deformations when the structure of interest is missing. In this work, we propose an empirical, greedy regularization term which promotes the tumor contraction. The proposed method is simple but very effective. It is based on a priori medical knowledge about the scar localization to promote the direction of the tumor propagation. The proposed method is compared to the Demons algorithm using both the artificially generated data with a known ground-truth and a real, medical data. A relative tumor volume reduction, a Hausdorff distance between the tumor beds, a RMSE between the deformation fields, and a visual inspection are used as the evaluation methods. The proposed method models the tumor resection accurately in the target data and improves the potential dose distribution for the radiotherapy planning.