Machine Learning for Modeling the Biomechanical Behavior of Human Soft Tissue
PURPOSE To investigate the feasibility of a biomechanical-based deformable image registration technique for the integration of multimodality imaging, image guided treatment, and response monitoring. METHODS AND MATERIALS A multiorgan deformable image registration technique based on finite element modeling (FEM) and surface projection alignment of selected regions of interest with biomechanical material and interface models has been developed. FEM also provides an inherent method for direct tracking specified regions through treatment and follow-up. RESULTS The technique was demonstrated on 5 liver cancer patients. Differences of up to 1 cm of motion were seen between the diaphragm and the tumor center of mass after deformable image registration of exhale and inhale CT scans. Spatial differences of 5 mm or more were observed for up to 86% of the surface of the defined tumor after deformable image registration of the computed tomography (CT) and magnetic resonance images. Up to 6.8 mm of motion was observed for the tumor after deformable image registration of the CT and cone-beam CT scan after rigid registration of the liver. Deformable registration of the CT to the follow-up CT allowed a more accurate assessment of tumor response. CONCLUSIONS This biomechanical-based deformable image registration technique incorporates classification, targeting, and monitoring of tumor and normal tissue using one methodology.