Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

Prostate cancer is the most common form of cancer and second leading cause of cancer death diagnosed in North American men, with more than 262,000 new cases and an estimated 33,660 deaths in 2013 [1, 2]. Given that the median patient survival time for metastatic prostate cancer ranges from 12.2 to 21.7 months [3], early diagnosis of clinically significant prostate cancer would have significant benefits to patient care. This is particularly true given that the 5 year survival rate after diagnosis for patients with prostate cancer at the non-metastatic stage is 96 % in Canada [4].

In the current clinical model, men with positive digital rectal exam (DRE) and elevated prostate-specific antigen (PSA) require multicore random biopsies for risk stratification. However, there is an ongoing controversy about the role of prostate PSA as a screening test in prostate cancer. Two recent major randomized clinical trials [5, 6] have demonstrated that PSA screening contains a significant risk of overdiagnosis for prostate cancer where it is estimated that 50 % of screened men are diagnosed with prostate cancer. This leads to painful needle biopsies and subsequent potential overtreatment [5, 6]. Moreover, it has become increasingly clear that prostate biopsies are harmful as they cause discomfort and possible sexual dysfunction and may result in increased hospital admission rates due to infectious complications [7]. Nevertheless, PSA testing has proven to reduce prostate cancer mortality by 20–30 % at long-term follow-ups [8]. Therefore, the PSA testing remains an important biomarker in diagnosing prostate cancers that are clinically significant. The remaining challenge is how to improve the prostate cancer diagnosis to reduce the overdiagnosis of clinically insignificant cancers.

Automatic detection of prostate cancer as part of a clinical decision support system can potentially help radiologists in interpreting images more accurately. Specifically, multi-parametric MR imaging (MP-MRI) which combines T2-weighted MRI (T2w), diffusion-weighted imaging (DWI), and dynamic contrast enhanced imaging (DCE) has been found to be a promising method for prostate cancer diagnosis and it has been used in different prostate cancer detection algorithms. By taking advantage of the unique information provided by each individual imaging technique, MP-MRI can exploit the different characteristics of prostate tissue to improve differentiation between cancerous and surrounding tissues. For example, cancerous tissue in the prostate gland may exhibit a moderate drop in signal in T2w [9] (which characterizes differences in transverse (spin-spin) relaxation time of tissue); restricted diffusion in DWI [9] (which characterizes diffusion of water in tissue); earlier onset time, higher peak, and shorter peak time in DCE [10] (which characterizes the concentration of an injected gadolinium contrast agent over time as it passes into the extracellular extravascular space of the tissue). Although DCE is considered as part of MP-MRI, T2w+DWI is the most common MP-MRI because it does not require invasive contrast agent as DCE does.

Radiologists’ interpretations of MP-MRI have shown to achieve good prostate cancer detection rates, reaching accuracies of 80 % in the peripheral zone of the prostate gland. Similarly, several algorithms have been proposed for auto-detection of prostate cancer using MP-MRI setting [11–14]. These algorithms usually compute a set of low-level features from the MP-MRI data to construct feature vectors. Next, a supervised classifier is trained using the computed feature vectors from the training cases and their associated ‘ground-truth’ labels (e.g., labeled healthy or cancerous). Finally, the trained classifier is used to classify new cases. The reported values for accuracy of cancerous versus healthy tissue classification ranges from 64 to 89 %, depending on the feature sets and training/test data.

The underlying challenge in all these auto-detection algorithms is whether there is enough separability between the cancerous and healthy tissues in the feature space. This means if the separability is poor, even sophisticated feature extraction algorithms may not have a significant effect on the accuracy of cancer detection. On the other hand, improving the separability of cancerous and healthy tissues in the images would have a significant impact on the performance of cancer auto-detection algorithms, potentially reducing the dependency on the feature extraction methods.

In this paper, we propose a new MP-MRI texture feature model that, in addition to T2w and conventional DWI, uses computed high-b diffusion-weighted imaging (CHB-DWI) [15] and the recently proposed correlated diffusion imaging (CDI) [16]. Compared to DWI, CHB-DWI and CDI have both shown initial promise to improve visual separability of cancerous and healthy tissues in prostate, which can lead to improved performance of the proposed MP-MRI texture feature model for detecting prostate cancer. To the authors’ best knowledge, the proposed texture feature model is the first that utilizes all of the above-mentioned MP-MRI modalities.

This paper is organized as follows: in Sect. 2, the proposed MP-MRI texture feature model is presented. Section 3 presents the testing methodology which includes the description of image data and the evaluation metrics used in this research. Sections 4 and 5 present the experimental results and conclusions, respectively.

2 Proposed Multi-Parametric Magnetic Resonance Imaging Texture Feature Model

In this section, we present the proposed MP-MRI texture feature model for prostate cancer and discuss the imaging and feature extraction methods used in the model.

2.1 Imaging Methods

This subsection summarizes the imaging methods used in the proposed MP-MRI feature model.

2.1.1 T2-Weighted Imaging

T2w is a MR imaging modality in which the sensitivity of tissue is characterized by measuring the relaxation time (spin-spin) of the applied magnetic field. The T2w imaging of the prostate usually shows a small reduction in signal in the cancerous tissue [9].

2.1.2 Diffusion-Weighted Imaging

DWI is a promising imaging modality in which the sensitivity of tissue to Brownian motion of water molecules is measured by applying pairs of opposing magnetic field gradient pulses [17]. The diffusion-weighted signal, S is measured as:

$$\displaystyle{ S = S_{0}e^{-bD} }$$
(1)

where S 0 is the signal intensity without the diffusion weighting, b consists of amplitude and duration of the diffusion pulses, and the time between the two pulses as well as the gyromagnetic ratio, and D represents the strength of the diffusion. The diffusion-weighted image (S) is usually generated with different b values which can be used to estimate D in Eq. 1, called apparent diffusion coefficient map (ADC), using least-squares or maximum likelihood strategies [18]. The cancerous tissue in ADC is usually represented by a darker intensity compared to the surrounding tissue.

2.1.3 Computed High-b Diffusion-Weighted Imaging (CHB-DWI)

Previous research has shown that high b-value DWI images (e.g., b-values greater than 1,000 s/mm2) allow for increased delineation between tumours and healthy tissues [15, 19] which makes the prostate cancer detection more robust. Nevertheless, due to hardware limitations, most MRI machines in practice do not produce DWI with b-values higher than 1,500 s/mm2. CHB-DWI is an alternative approach to obtain high-b DWI in which a computational model is used to reconstruct DWI at high b-values using low b-value DWI acquisitions [15]. For our experiments, we constructed CHB-DWI with b-value at 2,000 s/mm2 using the same least squares estimation technique used for ADC, extrapolating to the b-value of 2000 s/mm2.

2.1.4 Correlated Diffusion Imaging (CDI)

CDI [16] is a new diffusion magnetic resonance imaging modality, which takes advantage of the joint correlation in signal attenuation across multiple gradient pulse strengths and timings to not only reduce the dependency on the way diffusion gradient pulses are applied, but also improve delineation between cancerous and healthy tissue. The local correlation of signal attenuation across all b-values within a local sub-volume is calculated to better represent the overall characterization of the water diffusion properties of the tissue. The CDI signal is calculated as follows [16]:

$$\displaystyle\begin{array}{rcl} \mathit{CDI}(x)& =& \int \ldots \int _{b_{0}}^{b_{n} }S_{0}(x)\ldots S_{n}(x)P(S_{0}(x),\ldots,S_{n}(x)\vert V (x)) \\ & & \times dS_{0}(x)\ldots dS_{n}(x) {}\end{array}$$
(2)

where x denotes spatial location, S denotes the acquired signal, P denotes the conditional joint probability density function, and V (x) denotes the local subvolume around x.

2.2 Texture Feature Model

In order to separate the cancerous tissue from the healthy one, a set of features is calculated on a given MR imaging sequence (i.e., T2w, DWI, CHB-DWI, CDI). As part of the proposed MP-MRI texture feature model, we incorporate the texture features used in separate studies to establish a relationship between these features and tumour glucose metabolism and stage [20] and to predict the response of metastatic renal cell cancer to treatment [21]. These features include mean grey-level intensity (M), entropy (En), and uniformity (U), which are calculated as follows:

$$\displaystyle{ M = \frac{1} {N}\sum _{x,y=0}^{N-1}\mathit{pix}(i,j) }$$
(3)
$$\displaystyle{ \mathit{En} = -\sum _{l=1}^{k}p(l)\mathit{log}_{ 2}[p(l)] }$$
(4)
$$\displaystyle{ U =\sum _{ l=1}^{k}[p(l)]^{2} }$$
(5)

where pix(i, j) is the gray-level intensity in the pixel window, N is the window size, p is the probability density function of pixels in the window, and k is the number of grey levels in the image.

In addition, the proposed MP-MRI texture feature model incorporates another set of texture features extracted from the gray-level co-occurrence matrix (GLCM) in 4 directions: 0∘, 45∘, 90∘, and 135∘. The GLCM texture features are calculated as follows:

  1. 1.

    Contrast (Con)—a measure of the intensity difference between a pixel and its neighbors:

    $$\displaystyle{ \mathit{Con} =\sum _{ x,y=0}^{N-1}\vert x - y\vert ^{2}p_{\mathit{ glcm}}(x,y) }$$
    (6)
  2. 2.

    Energy (Eg)—the sum of squared elements in the GLCM;

    $$\displaystyle{ \mathit{Eg} =\sum _{ x,y=0}^{N-1}p_{\mathit{ glcm}}(x,y)^{2} }$$
    (7)
  3. 3.

    Homogeneity (H)—a value that measures the closeness of the distribution of elements in the GLCM to the GLCM diagonal.

    $$\displaystyle{ H =\sum _{ x,y=0}^{N-1} \frac{p_{\mathit{glcm}}(x,y)} {1 + (x - y)^{2}} }$$
    (8)
  4. 4.

    Correlation (Cor)—a measure of how correlated a pixel is to its neighbors:

    $$\displaystyle{ \mathit{Cor} =\sum _{ x,y=0}^{N-1}\frac{(x -\mu _{x})(y -\mu _{y})p_{\mathit{glcm}}(x,y)} {\sigma _{x}\sigma _{y}} }$$
    (9)

where p glcm is the probability value from the GLCM, μ x, μ y, σ x, and σ y are the means and standard deviation values of p glcm in horizontal and vertical directions, respectively.

As a result, the proposed MP-MRI texture feature model consists of a total of 19 features for each imaging modality: 16 from GLCM (4 in each direction) and 3 other texture features (i.e., mean gray level, entropy, and uniformity).

3 Testing Methodology

In the following, details about the image acquisition protocols and the performance measures are presented.

3.1 Image Data

MRI data of five patients were acquired using a Philips Achieva 3.0T machine at Sunnybrook Health Sciences Centre, Toronto, Ontario, Canada. All data was obtained under the local institutional research ethics board. For each patient, the following MP-MRI modalities were obtained (Table 1): T2w, DWI, and CDI. The patients’ age ranged from 53 to 75.

Table 1 Description of the prostate T2w, DWI, and CDI images
Full size table

3.2 Evaluation Metrics

We evaluated the performance of the proposed MP-MRI texture feature model for detecting prostate cancer via a leave-one-patient-out cross-validation using a Bayesian classifier. First, the feature extraction function was applied to each MR imaging sequence (i.e., T2w, ADC, CHB-DWI, and CDI) separately across all patients’ data. This generated four sets of features based on the proposed MP-MRI texture feature model. Next, a conventional MP-MRI setting was established by combining the features of T2w and ADC. As a new MP-MRI texture feature model, three configurations were realized: T2w+ADC+CHB-DWI, T2w+ADC+CDI, and CHB-DWI+ADC+CDI.

We used a Bayesian classifier to calculate sensitivity, specificity, and accuracy using leave-one-patient-out cross-validation approach. As ground-truth, all images were reviewed and marked as healthy and cancerous tissue by a radiologist with 18 and 13 years of experience interpreting body and prostate MRI, respectively.

4 Experimental Results

Table 2 shows sensitivity, specificity, and accuracy for all 8 MP-MRI modalities/models. For each modality, on average, 5,260 (5,110 healthy and 150 cancerous) and 1,315 (1,275 healthy and 40 cancerous) samples (i.e., pixel windows) were used as training and testing data, respectively for the leave-one-patient-out cross-validation. The results were averaged across all patients. In order to unify all three measures (sensitivity, specificity, and accuracy) for the purpose of comparison among different configurations, we also report the average of the three measures for each modality/configuration.

Table 2 Evaluation results for prostate cancer detection
Full size table

It is seen that CHB-DWI alone improves results compared to T2w and ADC (Sensitivity: 0.88 vs. 0.70 and 0.87, Specificity: 0.58 vs. 0.48 and 0.37, Accuracy: 0.58 vs. 0.49 and 0.38). Although CDI has a lower sensitivity compared to T2w (0.66 vs. 0.70), its specificity and accuracy is the highest among all individual modalities (0.80). This is due to the fact that CDI combines the information across b-values making it robust.

The first configuration of the new MP-MRI texture feature model, T2w+ADC+ CHB-DWI, improves the results with regard to the conventional model, T2w+ADC, (Sensitivity: 0.86 vs. 0.84, Specificity: 0.63 vs. 0.41, Accuracy: 0.64 vs. 0.42). The second configuration of the new MP-MRI texture feature model, T2w+ADC+CDI, loses the sensitivity compared to the conventional model, T2w+ADC, (0.68 vs. 0.84) but outperforms it in both specificity (0.81 vs. 0.41) and overall accuracy (0.81 vs. 0.42). The third configuration of the proposed MP-MRI (CHB-DWI+ADC+CDI) only loses 5 % in sensitivity compared to the conventional MP-MRI (T2w+ADC) (0.783 vs 0.836). In return, it improves the specificity and accuracy by 38 and 37 %, respectively (Specificity: 0.79 vs. 0.41, Accuracy: 0.79 vs. 0.42). In other words, with a slight reduction in true positive cases, a significant amount of false positive cases can be avoided. The best result for average of all three measures (i.e., sensitivity, specificity, and produced by CHB-DWI+ADC+CDI (0.79).

Figure 1 shows an example for all four modalities which include T2w, ADC, CHB-DWI, and CDI. As it can be seen, CDI (Fig. 1d) is the only modality that clearly shows a bright nodule where a tumour is located (confirmed by histopathology data).

Fig. 1
figure 1

(a) T2w does not clearly show a tumour although there is mild signal alteration in the left peripheral zone (arrow). (b) ADC does not clearly show a tumour (arrow). (c) CHB-DWI of 2,000 s/mm2 shows no tumour (arrow). (d) CDI clearly shows a bright nodule (arrow) corresponding to tumour

Full size image

5 Conclusions

In this paper, we introduced a new multi-parametric MRI texture feature model for prostate cancer detection. Our new MP-MRI texture feature model adds two new image modalities, CHB-DWI and CDI, to the most commonly used MP-MRI, T2w+ADC. We calculated a set of texture features for both the conventional MP-MRI and new MP-MRI texture feature models. A Bayesian classifier was trained via leave-one-patient-out setting to classify the new cases. We evaluated the proposed MP-MRI texture feature model in three configurations. The first configuration (T2w+ADC+CHB-DWI) improved the results (sensitivity, specificity, and accuracy) compared to the conventional MP-MRI. The second configuration (T2w+ADC+CDI) improved specificity (40 %) and accuracy significantly (39 %) with a loss in sensitivity (16 %) with respect to the conventional MP-MRI. The best result was achieved by the third configuration (CHB-DWI+ADC+CDI); it improved specificity and accuracy significantly (38 and 37 %, respectively) with a rather small loss in sensitivity (5 %) with respect to the conventional MP-MRI. The proposed MP-MRI texture feature model showed promise to tackle the overdiagnosis problem in prostate cancer detection.