MR PROSTATE SEGMENTATION VIA DISTRIBUTED DISCRIMINATIVE DICTIONARY (DDD) LEARNING

Abstract

Segmenting prostate from MR images is important yet challenging. Due to non-Gaussian distribution of prostate appearances in MR images, the popular active appearance model (AAM) has its limited performance. Although the newly developed sparse dictionary learning method[1, 2] can model the image appearance in a non-parametric fashion, the learned dictionaries still lack the discriminative power between prostate and non-prostate tissues, which is critical for accurate prostate segmentation. In this paper, we propose to integrate deformable model with a novel learning scheme, namely the Distributed Discriminative Dictionary (DDD) learning, which can capture image appearance in a non-parametric and discriminative fashion. In particular, three strategies are designed to boost the tissue discriminative power of DDD. First, minimum Redundancy Maximum Relevance (mRMR) feature selection is performed to constrain the dictionary learning in a discriminative feature space. Second, linear discriminant analysis (LDA) is employed to assemble residuals from different dictionaries for optimal separation between prostate and non-prostate tissues. Third, instead of learning the global dictionaries, we learn a set of local dictionaries for the local regions (each with small appearance variations) along prostate boundary, thus achieving better tissue differentiation locally. In the application stage, DDDs will provide the appearance cues to robustly drive the deformable model onto the prostate boundary. Experiments on 50 MR prostate images show that our method can yield a Dice Ratio of 88% compared to the manual segmentations, and have 7% improvement over the conventional AAM.

Keywords: Prostate segmentation; deformable segmentation; magnetic resonance image; sparse dictionary learning.

Fig. 1

Complicated non-Gaussian distribution of appearance features (gradient, intensity) in MR prostate images.

Fig. 2

Diagram of Distributed Discriminative Dictionary (DDD) learning framework. (a) A schematic explanation of distributed discriminative dictionaries, with each taking charge of tissue differentiation in a local region. (b) Diagram of training a discriminative dictionary, including key components of mRMR feature selection, sparse representation, and LDA learning.

Fig. 3

The partition of our deformable model. Sub-surfaces are indicated by different colors.

Fig. 4

ROC curves of tissue classification using different methods.

Fig. 5

Visual comparisons of segmentation results from 4 different methods. Red and yellow contours denote manual and automatic segmentations, respectively.