Weighted Partitioned Active Shape Model for Optic Pathway Segmentation in MRI

Active shape models (ASMs) have been established as robust model-based segmentation approaches and have been particularly relevant for objects ill-defined in image data. For example, the automatic segmentation of the optic pathway is almost impossible without shape models due to low contrast in MRI and local anatomical variability. However, traditional ASM is not optimal for complex or variable shapes segmentation due to its strong constraints. Herein, we introduce a weighted partitioned active shape model to improve the shape flexibility and robustness of ASMs and apply it to optic pathway (including the nerve, chiasm, and tract) segmentation. The strong constraints of ASM are relaxed by partitioning the whole shape into several subparts. In this way, the local shape variability can be captured and the number of training data can be reduced. Our novel weighted matching approach assigns a weight to each landmark point according to its appearance confidence, thus deforming the shape to reliable positions. In the application of optic pathway segmentation, the mean of root mean squared symmetric surface distance is 0.59 mm, which is about one voxel size.