Segmentation of prostate contours for automated diagnosis using ultrasound images: A survey

Highlights

• Survey focuses on Transrectal ultrasound (TRUS) images.

• Up-to-date comparative analysis of prostate segmentation approaches is proposed.

• Performance of different segmentation approaches are compared.

• Strength and limitation of various methods are discussed.

Abstract

Prostate cancer is the most common cancer that affects elderly men. The conventional non-imaging screening test for prostate cancer like prostate antigen (PSA) and digital rectal examination (DRE) tests generally lack specificity. Ultrasound is the most commonly available, inexpensive, non-invasive, and radiation-free imaging modality among all the screening imaging modalities available for prostate cancer diagnosis. The precise segmentation of prostate contours in ultrasound images is crucial in applications such as the exact placement of needles during biopsies, computing the prostate gland volume, and to localize the prostate cancer. Moreover, the low-dose-rate (LDR) brachytherapy treatment in which radioactive seeds are implanted in the prostate region requires accurate contouring of the prostate gland in ultrasound images. Therefore, it is very important to segment the prostrate region accurately for the diagnosis and treatment. This paper aims to present the analysis of existing approaches used for the segmentation of prostate in transrectal ultrasound (TRUS) images. In this survey, different segmentation methods used to extract the prostrate using criteria such as mean absolute distance, Hausdorff distance and time are discussed in detail and compared.

Introduction

After skin cancer, prostate cancer is the most common cancer in USA. According to American cancer society, there are about 180,890 new cases and 26,120 deaths due to prostate cancer in 2016 [1]. Prostate disease can be classified in to three categories: prostatitis, benign prostatic hyperplasia (BHP), and adenocarcinoma [2]. Prostatitis is an inflammation, usually caused by bacteria and can be cured by antibiotics. The BHP is due to the enlargement of prostate and requires surgery in the advanced stage. The adenocarcinoma is the cancerous state of the prostate and if not detected in an early stage, cancer spreads to other organs like bones, seminal vesicles and rectum. Hence, early detection of prostate cancer can significantly reduce the chances of death. The initial diagnosis of prostate cancer includes non-imaging tests such as digital rectal examination [3] and prostate specific antigen lack specificity [4]. These tests are used along with the medical imaging tests such as transrectal ultrasound (TRUS), magnetic resonance imaging (MRI) and computed tomography (CT) scan of prostate [5]. The variation in cellular density can be examined using MRI [6]. Increase in local perfusion can be studied with contrast enhanced computed tomography or MRI [12]. Ultrasound images are most commonly used for screening, as they are inexpensive in comparison with other medical imaging modalities, easy to use, and have no side effects. The ultrasound images are not inferior to MRI or CT images in terms of diagnostic importance [40].

Segmentation in ultrasound modality is highly dependent on the nature of data. Prostate ultrasound images contain high speckle noise, shadow artifacts due to calcification in prostate region, short range of gray levels and boundaries are missing especially at the base and apex region of the prostate [13], which makes the segmentation process difficult. However, ultrasound images are widely used in various computer aided diagnosis (CAD) systems [2], [7], [8], [9], [10], [11], image guided interventions, and various therapies like brachytherapy. In localized prostate cancer, low-dose-rate (LDR) brachytherapy can be used to provide a quick and accurate diagnosis [14]. The brachytherapy implants small radioactive seeds inside the prostate gland. But before the implantation, volume and shape of the prostate region is to be estimated. Therefore, it is very important to segment the prostate from TRUS image accurately. The shape and volume of the prostate is crucial for the diagnosis of prostate cancer [15]. In most of the clinical practice, manual or semi-automated segmentation is performed which is prone to observer variability due to the poor visibility of prostate in ultrasound images [16]. Moreover, the size and shape of the prostate is not fixed, as it changes because of bladder and rectum filling, cancer evolution rate and the effect of anesthesia. Hence, there is high requirement of the real time prostate delineation in clinical applications like brachytherapy where the decisions are made in the operating environment instantly before implantation of seeds [17].

Generally, ultrasound data is of two types (i) B-mode images and (ii) radio frequency (RF) signals. The B-mode ultrasound images are obtained by post-processing of RF signals captured from the transducer. Researchers have shown that there is significant amount of loss of information during this post-processing, which may be useful for classification of prostate region as cancerous or non-cancerous [18]. However, the conventional B-mode images are helpful in segmenting the prostate region. The texture-based and model-based procedures for prostate segmentation in TRUS images are depicted in Fig. 1, Fig. 2 respectively. As TRUS images have low signal-to-noise ratio, speckle noise, and a very large number of edge features which are not a part of prostate contour. Therefore, a pre-processing is required to smooth the image while preserving the edge information [64], [65], [66].

Feature extraction is a crucial step in segmentation process. The major challenge is to extract the most relevant features from the input dataset. Statistical analysis is needed to select the subset of relevant features in model construction. The selected feature sets are used to train the classifier with known training parameters. Model validation techniques like k-fold cross validation, leave-one-out cross validation can be used to validate the final outcome with respect to the ground truth available [68]. As depicted in Fig. 2, Model based segmentation can be either active contour or geometric deformable model based segmentation technique.

In this paper, the primary focus is on the techniques developed for prostate segmentation in TRUS images. The segmentation of prostate from TRUS is important in various stages of clinical decision making process. For example prostate gland segmentation is helpful in determining the volume of the prostate, which aids in brachytherapy and treatment of BHP. Moreover, the prostate boundary obtained from the segmentation process helps in image fusion with other modalities like magnetic resonance imaging and radio frequency data to localize the cancer. Manual contouring of prostate is a difficult task and prone to inter and intra observer variability. Hence, the computer aided semiautomatic or fully automated segmentation techniques are investigated in this paper.

Earlier, Ghose et al. [19] carried out a survey on segmentation of prostate in multiple modalities like TRUS, MR, and CT images. Nobel et al. [20] considered various types of tissues like breast, heart, and prostate in their survey and techniques for vascular disorder are also examined. Zhu et al. [21] published the survey not only focusing on the segmentation of prostate, but also on the classification of ultrasound and MR images in cancerous and non-cancerous, staging of cancer and registration of ultrasound and MR images. Hence the surveys [19], [20], [21] consider wider spectrum. Shao et al. [22] proposed a survey primarily based on the prostate region detection in ultrasound images. Although the survey reported good classification of segmentation techniques, but the precise comparison catalogue is missing for the better comparative analysis. Moreover, there has been significant improvement in the prostate segmentation techniques over the past decade with more focus on the development of fully-automated segmentation techniques which can be used in real time. Therefore, in this paper, we have provided an up-to-date, substantial categorization and comparative analysis of prostate segmentation approaches defined for B-mode TRUS images. Both 2D and 3D TRUS images are covered in this survey.

This paper is structured in five sections with Section 2 concisely representing the prostate ultrasound image enhancement, analysis, and challenges associated with them. The different classes of prostate segmentation approaches, brief description of related papers, Advantages and limitations of various model-based segmentation techniques are discussed in Section 3. The various validation techniques used for the segmentation are described in Section 4. The critical analysis and discussion of various methods are provided in Section 5. Finally, the conclusion and future trends are presented in Section 6.