Deformable Landmark-Free Active Appearance Models : Application to Segmentation of Multi-Institutional Prostate MRI Data

Prostate segmentation is a necessary step for computer aided diagnosis systems, volume estimation, and treatment planning. The use of standard datasets is vital for comparing different segmentation algorithms, and 100 datasets from 4 institutions were gather to test different algorithms on T2-weighted MR imagery. In this paper, a landmark-free Active Appearance Model based segmentation algorithm was employed to segment the prostate from MR images. A deformable registration framework was created to register a new image to a trained appearance model, which was subsequently applied to the prostate shape to yield a final segmentation. Results on 50 training studies yielded a median Dice coefficient of 0.80, and the fully automated algorithm was able to segment each prostate in under 3 minutes. 1 Background and Motivation Segmentation of the prostate boundary is useful for many applications, such as a computer aided cancer diagnosis systems [1], and treatment evaluation via planimetry-based volume estimation [2]. Several segmentation schemes for MR imagery of the prostate have been recently presented, including algorithms by Klein et al. [3], Martin et al. [4], Pasquier et al. [5], and Makni et al. [6]. Klein et al. [3] performed a registration between an MR image of the prostate and an atlas of training data to achieve a segmentation of the prostate. Martin et al. [4] also used an atlas of training images, but constrained the segmentation model through the use of a statistical shape model. Pasquier et al. [5] used an Active Shape Model [7] method for extracting a statistical shape model of the prostate, which then looked for strong gradients to identify the prostate edge. Finally, Makni et al. [6] used a statistical shape model of the prostate, and clustered the intensities within a manually placed region of interest into 3 clusters: surrounding tissues and fat, central prostate zone, and the peripheral prostate zone. Any pixels within the latter 2 zones were determined to be in the prostate. In this paper, we similarly perform image registration to align the prostate in a new image to a collection of existing training images, but do so within an Active Appearance Model (AAM) framework [8,9]. With AAM’s, Principal Component Analysis (PCA) is performed on the set of image intensities inside the object of interest, as well as a parameterization of the shape, to generate a low dimensional appearance projection. Traditionally, to segment a new image, the projections are varied, and the original, high dimensional shape and appearance are reconstructed. In the algorithm used in this paper (based on [9]), a new image is deformably registered to the AAM. At each iteration of the registration, the high dimensional set of intensities are reconstructed, and the best intensity reconstruction is found. This registration is then applied to the prostate shape, to yield a segmentation of the prostate. This algorithm is similar to the atlas-based approach [3], but instead of registering a new image to an atlas, the new image is registered to a statistical model of entire set of training data (defined using an AAM). While the original algorithm from which this is based [9] used an affine registration (scale, rotation, and translation) to register the new image to the AAM, this paper introduces a spline-based deformable registration component [10] to overcome differences in prostate shape and appearance due to highly varied multi-institutional data. These additional degrees of freedom are especially important when dealing with multi-institutional data (as in this dataset), because different acquisition protocols (such as the use, or lack of use, of the endorectal coil) can yield sub-par results when only linear (affine) transformations are considered. In addition, as in [9], a levelset representation of the shape to overcome issues associated with using landmarks such as triangulation errors and correspondence issues.