Statistical Intensity- and Shape-Modeling to Automate Cerebrovascular Segmentation from TOF-MRA Data

Complete, automatic, and fast segmentation of cerebrovascular Time-of-flight (TOF) Magnetic Resonance Angiography (MRA) is significant for the clinical application study, where vascular network coverage and accuracy are the focused issues. In our work, a novel statistical modeling method is proposed to efficiently improve the framework of Maximum a Posterior (MAP) and Markov Random Field (MRF), where the low-level process uses Gaussian mixture model to distinguish intensity information within the skull-stripped TOF-MRA data, and the high-level process embeds a new potential function of pair-wise sites into Markov neighborhood system (NBS). To explore vascular shape information in complex local context, the potential function employs vascular feature map and direction field. The Markov regularization parameter estimation is automated by using the machine-learning algorithm, which avoid the disadvantage of repeated trial-error-test to different data. This novel statistical model greatly improves segmentation accuracy and avoids vascular missing in the region of low contrast or low signal-to-noise ratio. Our experiments employ 109 public datasets from MIDAS data platform, in which 10 datasets is used to produce ground trues for quantitative validation. Existing statistical models are divided into 24 composite modes for cross comparisons, where the proposed strategy wins the best out of these modes on the evaluations.