Improved active shape models for segmentation of the prostate on MR imagery

Segmentation aims to determine which locations within an image contain the object of interest. Segmentation of the prostate boundary on clinical images is useful in a large number of applications including calculation of prostate volume pre- and post-treatment, detection of extra-capsular spread, and creation of patient-specific anatomical models. Manual segmentation of the prostate boundary is, however, time consuming and subject to inter- and intra-reader variability. T2-weighted (T2-w) Magnetic Resonance (MR) structural imaging (MRI) and MR Spectroscopy (MRS) have recently emerged as promising modalities for detection of prostate cancer in vivo. With the recent advance in prostate imagery, we have generated an accurate prostate segmentation system for MR imagery. Our system builds upon the popular Active Shape Model (ASM) framework, in which a statistical description of the shape is first generated, after which an appearance of the object of interest is modeled.;In our system, the shape model can be generated in either 2D or 3D, and is defined by a set of anatomical landmarks. For the appearance model, we offer several improvements. We generated statistical texture features of the prostate images, and use those features to overcome limitations of solely using intensities. In addition, we use intelligent feature selection algorithms including forward feature selection and adaboost to determine which features to include in our segmentation system. The statistical appearance models are not modeled as a simple Gaussian distribution, but rather as a sum of Gaussians, resulting in more accurate models. In 2D, a local appearance model is generated for each landmark location on the prostate border. However, in 3D this is infeasible, so we generate a global appearance model describing the voxels within the prostate. The 2D ASM resulted in a Dice similarity coefficient (DSC) of 0.85, while our 3D system resulted in a DSC of 0.89 (over 56 and 37 studies respectively). This is comparable to other state of the art prostate MR segmentation schemes. Finally, we have shown that in the specific application of prostate volume estimation, our system performs more accurate volume estimations than currently employed clinical models. Our system achieved a correlation (R2 value) with the ground truth volume of 0.82 while the clinical model achieved an R2 value of 0.70. Our system had a volume fraction of 1.05 in comparison to the ground truth volume, while the clinical model achieved a volume fraction of 1.14. Overall, we have developed an efficient, accurate, and useful prostate segmentation scheme for MR imagery.