A novel class of machine-learning-driven real-time 2D/3D tracking methods:texture model registration (TMR)

摘要

We present a novel view on 2D/3D image registration by introducing a generic algorithmic framework that is based on supervised machine learning (SML). First and foremost, this class of algorithms, referred to as texture model registration (TMR), aims at making 2D/3D registration applicable for time-critical image guided medical procedures. TMR methods are two-stage. In a first offline pre-computational stage, a prediction rule is derived from a pre-interventional 3D image and according geometric constraints. This is achieved by computing digitally reconstructed radiographs, pre-processing them, extracting their texture, and applying SML methods. In a second online stage, the inferred rule is used for predicting the spatial rigid transformation of unseen intra-interventional 2D images. A first simple concrete TMR implementation, referred to as TMR-PCR, is introduced. This approach involves principal component regression (PCR) and simple intermediate pre-processing steps. Using TMR-PCR, first experimental results on five clinical IGRT 3D data sets and synthetic intra-interventional images are presented. The implementation showed an average registration rate of 48 Hz over 40000 registrations, and succeeded in the majority of cases with a mean target registration error smaller than 2 mm. Finally, the potential and characteristics of the proposed methodical framework are discussed.