Automatic SWI Venography Segmentation Using Conditional Random Fields
The human cerebrovasculature is extremely complicated and its three dimensional (3D) highly parcellated models, though necessary, are unavailable. We constructed a digital cerebrovascular model from a high resolution, 3T 3D time-of-flight magnetic resonance angiography scan. This model contains the arterial and venous systems and is 3D, geometric, highly parcellated, fully segmented, and completely labeled with name, diameter, and variants. Our approach replaces the tedious and time consuming process of checking and correcting automatic segmentation results done at 2D image level with an aggregate and faster process at 3D model level. The creation of the vascular model required vessel pre-segmentation, centerline extraction, vascular segments connection, centerline smoothing, vessel surface construction, vessel grouping, tracking, editing, labeling, setting diameter, and checking correctness and completeness. For comparison, the same scan was segmented automatically with 59.8% sensitivity and only 16.5% of vessels smaller than 1 pixel size were extracted. To check and correct this automatic segmentation requires 8 weeks. Conversely, the speedup of our approach (the number of 2D segmented areas/the number of 3D vascular segments) is 34. This cerebrovascular model can serve as a reference framework in clinical, research, and educational applications. The wealth of information aggregated with its quantification capabilities can augment or replace numerous textbook chapters. Five applications of the vascular model were described. The model is easily extendable in content, parcellation, and labeling, and the proposed approach is applicable for building a whole body vascular system.