Ground Truth Estimation using Multimodal Data Fusion

The accurate localization of prostate cancer is important as it can affect both surgical and radiation treatment planning. Generating accurate ground truth is critical for reliable validation. Although, some part of this ground truth is contained in certain set of data such as pathological analysis and images of the organ, there exists some degree of random error from inter/intra-expert variabilities and surgical manipulation. Depending on the imaging method used, the cancer sites would have displaced according to the surface deformation and elasticity of the gland. We are proposing a new method to eliminate the random error in data acquired from different techniques and move closer to the ground truth of the data using multimodality dataset fusion. A non rigid registration method is used to register and fuse the volume data obtained from transrectal-ultrasound, magnetic resonance imaging and laser scan volume obtained from the same patient. The proposed registration method is based on a combination of iterative closest point (ICP) algorithm and multilevel B-splines. This is followed by a 3-D reconstruction to rescale the fused image; we will study and compare the results from two methods, (i) linear rescaling using Prolate Ellipsoid Equation and (ii) non-linear scaling using Generalized Hough Transform and prostate shape equations to estimate the ground truth of the model. This should provide an estimation of the random error in each image modality. Localization of prostate cancer in each modality is expected to be improved after the removal of the random error.