Prostate Cancer Segmentation Using Multispectral Random Walks

Several studies have shown the advantages of multispectral magnetic resonance imaging (MRI) as a noninvasive imaging technique for prostate cancer localization. However, a large proportion of these studies are with human readers. There is a significant inter and intra-observer variability for human readers, and it is substantially difficult for humans to analyze the large dataset of multispectral MRI. To solve these problems a few studies were proposed for fully automated cancer localization in the past. However, fully automated methods are highly sensitive to parameter selection and often may not produce desirable segmentation results. In this paper, we present a semi-supervised segmentation algorithm by extending a graph based semi-supervised random walker algorithm to perform prostate cancer segmentation with multi-spectral MRI. Unlike classical random walker which can be applied only to dataset of single type of MRI, we develop a new method that can be applied to multispectral images. We prove the effectiveness of the proposed method by presenting the qualitative and quantitative results of multispectral MRI datasets acquired from 10 biopsy-confirmed cancer patients. Our results demonstrate that the multispectral MRI noticeably increases the sensitivity and jakkard measures of prostate cancer localization compared to single MR images; 0.71 sensitivity and 0.56 jakkard for multispectral images compared to 0.51 sensitivity and 0.44 jakkard for single MR image based segmentation.