计算机科学与系统生物学杂志

A Novel NMF Guided Level-set for DWI Prostate Segmentation

Abstract

Patrick McClure, Fahmi Khalifa, Ahmed Soliman, Mohamed Abou El-Ghar, Georgy Gimelfarb, Adel Elmagraby and Ayman El-Baz

Objective: To develop an automated 3D framework for prostate segmentation from diffusion-weighted imaging (DWI).

Methods: The proposed framework integrates level-set deformable model and nonnegative matrix factorization (NMF) techniques. In the proposed framework, the level-set is guided by a novel speed function that is derived using NMF, which extracts meaningful features from a large dimensional feature space. The NMF attributes are calculated using information from the DWI intensity, a probabilistic shape model, and the spatial interactions between prostate voxels. The shape model is constructed using a set of training prostate volumes and is then updated during the segmentation process using an appearance-based method that takes into account both a voxel’s location and its intensity value. The spatial interactions are modeled using a second-order pairwise 3D Markov-Gibbs random field (MGRF).

Results: We tested our framework on 125 in vivo DWI-MRI prostate data sets that has been collected from 10 subjects at b-values ranging from 0 to 1000 s/mm2. Our results show that using NMF-based feature fusion to guide the level-set increases the segmentation accuracy compared with previously proposed methods using two evaluation metrics, the dice similarity coefficient (DSC) and Hausdorff distance (HD). The proposed method achieved an average DSC of 0.868 ± 0.03 and an average HD of 5.61 ± 2.12 mm compared to an average DSC of 0.83±0.07 and an average HD of 6.67 ± 1.84 mm for a maximum a posteriori (MAP)-based level-set method, and an average DSC of 0.810 ± 0.05 and an average HD of 9.07 ± 1.64 for a level-set driven only by intensity and shape information.

Conclusions: Experimental results show that the integration of 3D intensity, shape, and spatial features with NMF-based feature fusion increases the ability of the proposed method to perform robust prostate segmentation despite image noise, inter-patient anatomical differences, and the similar intensities of the prostate and surrounding tissues.