Prostate Ultrasound Image Processing

Document Sample
Prostate Ultrasound Image Processing Powered By Docstoc
					     Deian Stefan

        // Creating a hardware accelerated image with variable                     References
        // alpha transparencies.                                                   1. Clingman, K. M. 2004. Practical Java Game Programming. Charles
        BufferedImage imageAlpha = graphicConf.                                       River Media.
                                                                                   2. Denault, A. 2005. Minueto, an Undergraduate Teaching Develop-
                                                                                      ment Framework. Master’s thesis, McGill University (Aug.).
     Rendering to the Screen                                                       3. Haase, C. 2003. BufferedImage as good as butter part 1. http://
     As previously mentioned, images should first be drawn on the memory     (accessed 14 Aug. 2003).
                                                                                   4. Haase, C. 2004. Toolkit buffered volatile managed image strategies.
        // Drawing an image on the memory buffer                             (accessed 11 Aug. 2004).
        graph2D.drawImage(bufferedImage, x, y, null);
                                                                                   5. Java gaming forums.

        Once the frame is ready, we can render the content of the memory           6. Minueto.
     buffer to the screen.
        // Rendering the current buffer to the screen                              Alexandre Denault ( is a PhD student
        graph2D.dispose();                                                         in the School of Computer Science at McGill University. His research;                                                     interests include fault tolerant distributed environments, teaching tools,
        graph2D = (Graphics2D)bufferStrategy.getDrawGraphics();                    and game development (specifically content generation and cheat preven-
                                                                                   tion). He is an active member of both the Mammoth research project and
     Conclusion and Future Work                                                    CSGames, an inter-university computer science competition.
     Despite our initial apprehension about using Java for game develop-
     ment, Java 2D has shown to be very efficient, if used properly. Two           Jörg Kienzle ( is an assistant professor at the
     years of using Minueto for the game development of COMP-361 proj-             School of Computer Science at McGill Universiy in Montreal, Canada,
     ects, the systems development project course at McGill, has shown             where he is leading the Software Engineering Laboratory. He holds a PhD
     that it performs extremely well, thanks to the different optimizations        and engineering diploma from the Swiss Federal Institute of Technology in
     presented in this article. Our future work in this area will focus on effi-   Lausanne (EPFL). His current research interests include fault tolerance,
     cient ways to combine Swing and Java 2D technology.                           software development methods, aspect-orientation, and computer games.

                                  Prostate Ultrasound
                                  Image Processing
                                  by Deian Stefan

            rostate cancer is the second most common cancer found in men in the United States and
            a leading factor in cancer-related deaths. It is a major health issue, causing a dramatic
            impact on the lifestyle of elderly men. Treatment is available, but it is most successful
     when the cancer is in an early stage. Treatment of an advanced cancer, especially if it has spread
     beyond the prostate, can be very complex and in many cases incurable. To find the cancer still
     in an early stage, it is important for men over the age of 50 to get an annual screening. Presently,
     several different screening techniques are used to diagnose the presence of the prostate carci-
     noma, the most common of which are digital rectal examination (DRE), prostate specific anti-
     gen (PSA) blood test, and transrectal ultrasound (TRUS). In cases where the presence of cancer
     is highly suspected, a biopsy usually follows the screening to fully confirm the diagnosis.
20    Spring 2007/ Vol. 13, No. 3                                                                               Crossroads
                                                                                                           Prostate Ultrasound Image Processing

    The DRE is a procedure in which the doctor checks the size,                 Speckle Reduction
shape, and texture of the prostate by inserting a finger into the rectum        Traditional Filtering
of the patient. This procedure is not the most accurate test available          Low-pass filters are successful in filtering the speckle from the TRUS
because only the back wall of the prostate is examined, while the mid-          prostate images, because speckle has similar characteristics to random
dle and front of the prostate remain undiagnosed. In addition, only if          multiplicative noise. Low-pass filters filter out such noise by attenuat-
the cancer is in an already advanced stage will a noticeable difference         ing high frequency detail in the image based on global or regional
in the texture be recognized.                                                   details. The simplest and most commonly used low-pass filter is the
    The PSA test, which commonly accompanies the DRE, measures                  arithmetic mean filter, or averaging filter. The averaging filter is math-
the amount of prostate specific antigen in the blood. A high PSA level          ematically defined as:
usually indicates the presence of cancer; however, because common                   Filter 1 (Arithmetic Mean Filter): Given original image f and a
diseases and infections of the prostate, such as benign prostatic hyper-        small neighborhood of n pixels, Sx y, surrounding pixel (x,y) such that
                                                                                Sx y f and (x,y) Sx y, the gray level of pixel (x,y) in filtered image
trophy (BPH) and prostatitis, cause an elevation in the PSA level, the
                                                                                g is g(x,y) = 1⁄n       Sx y(i), where Sx y(i) is the ith of n pixels in Sx y.
accuracy of the diagnosis is indeterminable. Furthermore, patients with
                                                                                    From the above definition, one can see that if a pixel is part of the
normal PSA levels could still be diagnosed with cancer at a later time.
                                                                                noise, its gray level is attenuated by taking the average gray level of the
    TRUS is the most common medical imaging modality used in
                                                                                pixels surrounding it. The averaging filter smooths areas of high fre-
prostate cancer screenings, followed by computerized axial tomogra-
                                                                                quency and only minimally changes already constant areas. A more
phy (CAT), which is used to diagnose the spread of the cancer beyond
                                                                                complicated filter is the median filter, which is defined as:
the prostate. TRUS is inexpensive, provides the radiologist with                    Filter 2 (Median Filter): Given the details of Filter 1, the gray level
enough detail to recognize any abnormalities within the prostate, and           of pixel (x,y) in g is g(x,y) = median{Sx y(i)}, for i = 1… n.
can also be used in guiding a real-time biopsy. TRUS usually accom-                 The above non-linear filter is more robust than the averaging filter.
panies the DRE and PSA test, because it does not provide enough                 Gray levels that are likely to be noise will not be averaged into the
detail to be used as the only screening technique in the diagnosis.             interpolated result, because the filter chooses the gray level that is
    In a typical TRUS screening, the prostate is manually delineated in         closest to the majority of the pixels?the median gray level. Similar to
the ultrasound image in order to calculate its size and volume, which           the above filters in taking into account the values of neighboring pix-
are additional details used to support the diagnosis. Because manual            els to interpolate the new pixel value, although with better perform-
segmentation of the prostate is very tedious and usually prolongs the           ance than the arithmetic mean filter in keeping edge detail, is the
diagnosis, automated delineation of the prostate is preferable.                 geometric mean filter defined below:
Computer aided segmentation leads to faster, more accurate, and more                Filter 3 (Geometric Mean Filter): Given the details of Filter 1, the
precise results without much interference from the doctor [8].                  gray level of pixel (x,y) in g is g(x,y) =       Sx y(i) µ, where µ = 1⁄n .
Segmentation of the prostate within the ultrasound image can be fur-                Filters 1 through 3 replace the gray level f (x,y) by taking into
ther used for real-time biopsy guidance, automatic volume determina-            account the surrounding detail and attenuating the noise by lowering
tion, and computer automated analysis of the prostate for the                   the variance, 2. These filters are known as smoothing spatial filters,
detection of cancer or abnormal regions.                                        with the median and geometric mean filters outperforming the arith-
    Different methods of delineating the prostate, such as those presented      metic mean filter in reducing noise while still preserving edge details.
in [3, 6, 8, 9], have shown to be successful in segmenting the prostate.        Increasing neighborhood size, n, results in higher noise attenuation,
However, they still require the input of the doctor to identify the prostate.   but also loss of edge detail. For further details on spatial filters and in-
The present study presents an algorithm for automatically identifying the       depth explanations, see [4].
prostate within the TRUS image. Successful automatic identification of
the prostate is crucial to the success of an automatic segmentation algo-       Sticks Filtering
                                                                                Further study of the performance of low-pass filters and adaptive low-
rithm. A method of delineating the prostate is also presented.
                                                                                pass filters has shown that while these filters successfully remove
                                                                                much of the speckle, they can cause a loss of detail in low-contrast
Algorithm Design
                                                                                border regions. An alternative method of filtering speckle using sticks,
                                                                                as proposed in [1] and utilized in [8], as a smoothing filter was imple-
Because region boundaries in medical ultrasound images are linear
                                                                                mented with success.
features created by different tissue properties, detection of the prostate
                                                                                    The sticks filtering algorithm takes on the challenge of filtering
body in a TRUS image can be understood as detection of edges. The               speckle in ultrasound images, without losing edge detail, by determining
detection of edges in ultrasound images is a challenging task because           whether a linear feature passes through pixel (x,y) and then calculating
of the high level of speckle noise corrupting the image. High contrast          the filtered pixel intensity g(x,y), which is the arithmetic mean of neigh-
regions can, however, be successfully detected by low-pass filtering the        boring pixels in the direction of the stick—the most likely direction of
image and applying edge-detection operators (e.g. Sobel, Prewitt, or            the linear feature passing through (x,y). Below is a formal definition.
Laplacian of Gaussian [5]) to highlight the linear features.                        Filter 4 (Sticks Filter): Given original image f, stick length n, and
    TRUS prostate images present the additional difficulty of low con-          stick thickness k, the set of all sticks is S = s i i = 1… 2n–2 , where
trast boundaries, which are difficult to identify after low-pass filtering      s i is a stick of length n, thickness k, and orientation i. Assuming n is
because low-pass filters, while creating a reduction in speckle noise,          the length in pixels, there are 2n–2 possible orientations the stick can
also reduce the amount of detail in the image. A different filtering            be uniquely arranged in. Mathematically, a stick can be described as a
method, presented by Czerwinski in [1], is used for speckle filtering.          spatial filtering mask:

Crossroads                                                                                    Spring 2007/ Vol. 13, No. 3        21
     Deian Stefan

        s i (s,t) = 1⁄n , if (s,t) is along angle i ; 0 otherwise ,
                   for all (s,t) D s i                                         (1)

        From the above definition of a stick, it is important to see that con-
     volving f with s i smooths the speckle while highlighting the linear
     features in direction i . However, pixels with lines passing through
     them in direction j, where j i, will be assigned an undesired gray
     level. To correctly filter f, it is important to define the set of all images,
     each highlighting a different direction i :
        H = hi i = 1… 2n–2 , where hi = f * s i                                (2)
        Each pixel (x,y) will then have a gray level g(x,y), such that,
        g(x,y) = max hi(x,y) , for i… 2n–2                                     (3)

         By implementing the concept of a stick passing through each pixel
     (x,y) and using that stick as a basis for the interpolated intensity g(x,y),
     the final filtered image g will be smooth and contain contrast
     enhanced region borders. An example of a set of sticks S, of length five
     pixels and one pixel thickness is shown in Figure 1, where the white
     and black pixels have a value of 0 and 1⁄5, respectively.
                                                                                         Figure 2: a) Original image; b) Filtered: n = 5 k = 1; c)
                                                                                         Filtered: n = 7 k = 3; d) Filtered: n = 15 k = 1.

                                                                                      fs. Similar to the algorithm presented in [9], the top-hat transformation
                                                                                      hs and bottom-hat transformation bs of fs are used to increase the con-
                                                                                      trast near the edges of the image according to the following equation:
                                                                                         fc = f + hs – bs                                                   (4)

                                                                                         Small regions in fc are then filled, resulting in ff , which is further
                                                                                      transformed by a global thresholding function, to get ft, where:
                                                                                         ft(x,y) = 1 + (m ⁄ ff (x,y)) E –1,
     Figure 1: Set of sticks with n = 5 and k = 1.                                                 where m = max ff – mean ff and E = 7                     (5)

         Various stick lengths and thicknesses have different filtering effects.
                                                                                          The thresholded image is then binary thresholded, and a dilation
     Increasing stick length leads to a more smoothly filtered image, at the
                                                                                      with a small disk is performed to close any small gaps, with fb as the
     expense of weakly highlighting tightly bound curves—a result of the stick
                                                                                      result. The binary image fb is then eroded k–1 times with a similar
     being longer than some of the boundary edges. Similarly, thicker sticks
                                                                                      small disk used for the dilation, assuming that the kth erosion will
     suppress more noise, at the expense of making thin boundaries less vis-
                                                                                      result in an empty image. The result is a small number of pixels within
     ible. A thick stick can be used to smooth noise, similar to an arithmetic
                                                                                      the prostate body. Please see [4] for a description of thresholding func-
     mean filter, with the addition of highlighting broad region differences.
                                                                                      tions, including (6), and further information on morphological opera-
         Generally, it is important to implement a stick length that is longer
     than the noise correlation length but shorter than the length over which         tors (e.g., erosion, dilation, top-hat, bottom-hat, etc.).
     the boundary edges are straight. Figure 2 shows an image filtered with dif-          The performances of both the algorithm proposed in [9] and the
     ferent stick lengths and thicknesses. All three filtered images clearly show     algorithm proposed in the present study were tested on 144 random
     an increase in contrast near the prostate borders and reduced speckle            TRUS images. Figure 3 shows both a successful and unsuccessful
     noise with increasing stick length. An in-depth study of the effects of stick    mapping of the prostate. The algorithm proposed in the present study
     length and thickness on ultrasound image filtering is presented in [2].          had an 88.9% success rate in mapping a point in the prostate body,
                                                                                      while the algorithm proposed in [9] had an 84.0% success rate, giving
     Region Identification                                                            the present algorithm an improvement of 4.9%.
     The algorithm used in the present study to identify the prostate body
     region is based on the prostate center seeking algorithm presented in [9].       Region Delineation
         Algorithm 1 (Region Identification): Given original image f, the             Given the mapped point within the prostate body, the second goal of
     speckle noise is low-pass filtered using the sticks filter with a stick          the present study is to detect the edges of the prostate, moving radially
     length of five pixels and thickness of three pixels to get fl, which is again    out from the found point. To highlight the edges, the image is filtered
     filtered using the sticks filter; however, now with a stick length of fif-       to reduce speckle noise, and different traditional operators, including
     teen pixels and thickness of one pixel in order to get the smooth image          Sobel, Canny, Prewitt, and Laplacian of Gaussian are used to create

22    Spring 2007/ Vol. 13, No. 3                                                                                  Crossroads
                                                                                                           Prostate Ultrasound Image Processing

                                                                                ary, an algorithm based on sticks filtering to detect the body of the prostate
                                                                                is developed, with a success rate of 88.9%. The second goal of the study,
                                                                                to delineate the prostate body, was unsuccessful. The GVF snake failed to
                                                                                even closely converge toward the prostate boundaries. Studies of incor-
                                                                                porating different prostate segmentation methods are presented in [3, 6,
                                                                                8, 9] and will be considered as alternatives in future studies.

                                                                                The author would like to thank Prof. Hong Man for his discussions con-
                                                                                cerning this work, Rafa Llobet [7] for providing the case images, Prof. Yu-
 Figure 3: a) Successful mapping; b) Unsuccessful mapping.                      Dong Yao for administrating the program, and the National Science
                                                                                Foundation for sponsoring the study.
an edge map. The edge map is then used to create an external force
for active contours, called the gradient vector flow (GVF) [11].                References
    An active contour, or snake, is a curve, x(s), with internal (tension       1. Czerwinski, R. N., Jones, D. L. and O’Brien, Jr., W. D. 1998. Line
and rigidity) and external (linear features and boundaries) forces, that           and boundary detection in speckle images. IEEE Trans. on Image
moves through the spatial domain of an image, minimizing an energy                 Process. 7, 12 (Dec.) 1700-1714.
function, E, where
                                                                                2. Czerwinski, R. N., Jones, D. L. and O’Brien, Jr., W. D. 1999. Detec-
   E=      1⁄ α x (s) 2 + β x (s) 2 + E x(s) ds,                         (6)
            2                          ext                                         tion of lines and boundaries in speckle images: Application to med-
and α is the adjustable weight of the snake’s tension, β is the adjustable         ical ultrasound. IEEE Trans. on Medical Imag. 18, 2 (Feb.) 126-136.
weight of the snake’s rigidity, and x and x are the first and second deriv-
                                                                                3. Fitzpatrick, J. M. and Reinhardt, J. M., Eds. 2005. Prostate ultra-
atives of x, with respect to s. A normalized GVF field is used as the exter-
                                                                                   sound image segmentation using level set-based region flow with
nal force to minimize E and iteratively solve the GVF snake according to:
                                                                                   shape guidance. SPIE (April).
   xt (s,t) = α x t (s,t) – β x t (s,t) + v,
                                                                                4. Gonzalez, R. C. and Woods, R. E. 2002. Digital Image Processing
              where t is time and v is the GVF field                     (7)
                                                                                   2nd. Ed. Prentice-Hall.
The above equations and further details regarding the solution to the
                                                                                5. Gonzalez, R. C., Woods, R. E. and Eddins, S. L. 2004. Digital Image
GVF snake are explained in detail in [10] and [11].
                                                                                   Processing Using MATLAB®. Pearson, 395-407.

                                                                                6. Jendoubi, A., Zeng, J. and Chouikha, M. F. 2004. Top-down
                                                                                   approach to segmentation of prostate boundaries in ultrasound
                                                                                   images. In AIPR. 145-149.

                                                                                7. Llobet, R., Perez-Cortes, J. C., Toselli, A. H. and Juan, A. 2006.
                                                                                   Computer-aided detection of prostate cancer. Int. J. Med. Informatics
                                                                                   (To appear).

                                                                                8. Pathak, S. D., Chalana, V., Haynor, D. R. and Kim, Y. 2000. Edge-
                                                                                   guided boundary delineation in prostate ultrasound images. IEEE
                                                                                   Trans. Med. Imag. 19, 12. 1211-1219.
           Figure 4: The initial snake (yellow) converged
            (red) toward the prostate boundary (green).                         9. Sahba, F., Tizhoosh, H. R. and Salama, M. M. A. 2005. Segmentation
                                                                                   of prostate boundaries using regional contrast enhancement. In IEEE
    The GVF snake is successful in delineating objects in images with low          International Conference on Image Processing (ICIP) vol. 2 (Sept.).
speckle, including MRI images. However, Figure 4 shows the result of a             1226-1269.
GVF snake (red) in an ultrasound image, initialized on the point detected       10. Xu, C. and J. L. Prince, J. L. 1998. Generalized gradient vector flow
within the prostate (yellow). To solve the GVF, the original image was fil-         external forces for active contours. Signal Process.: An Int. J. 71, 2
tered using the sticks filter, and the Canny method was used to create              (Dec.). 131-139.
the edge map to which the GVF algorithm, with 500 iterations, was then
applied. As shown in Figure 4 the convergence of the GVF snake toward           11. Xu, C. and Prince, J. L. 1998. Snakes, shapes, and gradient vector
the prostate boundary (green) clearly failed, as can be seen by the pres-           flow. IEEE Trans. Image Process. 7, 3 (March). 359-369.
ence of false boundaries within the prostate after filtering, indicating the
need to improve speckle filters and edge detection methods.                     Biography
                                                                                Deian Stefan ( is a second year undergraduate elec-
Conclusion                                                                      trical engineering student at The Cooper Union. His research interests
The present study introduces a different method of filtering speckle noise      include security, operating systems, signal processing, image processing,
without a significant loss of edge detail in the filtering process. Under the   and mathematics. He is an active member of the ACM and the IEEE.
assumption that the prostate body has a lower average gray than its bound-      Website:

Crossroads                                                                                    Spring 2007/ Vol. 13, No. 3        23

Shared By: