8
1 1

A speckle suppression method for medical ultrasound images based on local statistics

Embed Size (px)

Citation preview

1

1

2

A Speckle Suppression Method for MedicalUltrasound Images Based on Local Statistics

Gjenna Stippel, Ivana Duskunovic, Alexandra Pizurica, Filip Rooms, Wilfried Philips, IgnaceLemahieu

Abstract— White Matter Damage (leukomalacia)is detected by the visual inspection of ultrasoundimages of the neonatal brain by an expert. A needfor a (semi) computerised way of segmenting the in-fected regions, in order to make quantitative mea-surements as an aid to the subjective diagnosis, isfelt. The performance of snakes for this purpose,however, is heavily deteriorated by the presence ofspeckle noise.

In this paper a new filter, based on local statisticsin the image, is proposed, that removes a significantamount of speckle noise in the healthy parts, whileit makes the areas infected by WMD more uni-form, thus severely improving the performance ofthe snake. The results after applying the proposedtechnique are compared with the manual segmen-tation done by an expert.

Keywords— speckle, medical ultrasound, texture,adaptive filter, segmentation, snake

I. introduction

20 Percent to 50 percent of the very low birth weightinfants (VLBW: <1500g), suffer from “White Mat-ter Damage” (leukomalacia). Generally WMD is vis-ible in an ultrasound image as a “zone of increasedechodensity” (i.e. a typical kind of “white cloud”),called a “flare”. See, for example, the figures 1,2, and3.

Nowadays, the diagnosis of WMD is still solely de-pendent on the subjective visual interpretation of anexpert. A need for a more quantitative method as anaid for making a correct diagnosis is felt. In order tomake any measurements on these flares (semi) com-puterised, we should find a good segmentation first.Using the “snakes” technique for this purpose is a clas-sical approach. However, applying a snake on the un-processed image does not give satisfying results, be-

Ghent UniversityDepartments TELIN and ELISSint-Pietersnieuwstraat 41B-9000 Ghent-BelgiumTel.: +32 9 264 34 26Fax.: +32 9 264 42 [email protected]

Fig. 1. Healthy brain

Fig. 2. Ill brain

cause the snake “gets stuck” in the speckle noise inthe healthy tissue. In this paper we present an adap-tive filter which suppresses the speckle noise in thehealthy tissue, while, at the same time, it enhancesthe uniformity in the sick tissue, thus increasing theperformance of the snake. The results are comparedwith the manually drawn segmentations of an expert.

II. parameters

We investigated 48 ultrasound images, all of whichwere classified by a neonatologist as either certainly

3

Fig. 3. Infected region

ill (i.e. suffering from WMD), or certainly healthy;16 of these images were from a healthy child, and32 from a child suffering from WMD. First we theimages were processed using the compensation algo-rithm described in [1]. In all of the images we selecteda rectangle of 32x30 pixels at exactly the same spot.(See fig 4). According to the neonatologist, there willcertainly be a zone of increased echodensity on thatplace, if the child suffers from WMD. Within the rect-angle we calculated several parameters, under whichthe mean grey-value and the “contrast” (in section IVwe describe exactly what we mean by the contrast of aregion). These last two turned out to be distinctive indetermining whether it was an ill or a healthy area. Ascatter plot of the result is shown in figure 5. The sep-arate cluster in the left bottom corner indicates thata mean grey-value of less then 67, and a contrast ofless then 35 mean that the tissue is healthy, otherwiseit is ill. Similar results, but for ultrasound images ofthe prostate, were presented in [2]. The complete re-search is still in progress; we used these preliminaryresults to construct the proposed filter.

III. snake

The snake we used for our research is the GVFsnake, as presented in [6]end [7]. In the figures 6,7, 8 and 9 the original image, the delineation of theinfected region by the expert, the initial snake posi-tion, and the final boundary which the snake foundare shown. As can be seen from this example, theperformance of the snake is heavily deteriorated bythe speckle noise in the healthy tissue.

Fig. 4. The specific rectangle (simulation)

Fig. 5. Results of measurements

Fig. 6. Original image

4

Fig. 7. Deliniation by expert

Fig. 8. Initial snake position

Fig. 9. Final snake position

compensation

image

image

originalimage

filteredimage

algorithm

growing

region

filter

median

procedure

lighting ordarkening

compensated

levelled

initialsegmentation

parametersmean grey-value contrast

Fig. 10. The structure of the filter

IV. filter

In order to improve the performance of the snake,we now describe a filter, which suppresses the specklein the healthy tissue, while at the same time it in-creases the uniformity in the ill tissue.

In figure 10 the overall procedure is shown. As isindicated in it, the procedure of the filter can be de-scribed in five steps:

1) First, out of the original image a “compensated im-age” is constructed, with the compensation algorithmintroduced in [1].2) Then, by applying a median filter to this com-pensated imaged, a “blurred compensated image” ismade.3) In this blurred image an initial segmentation ismade using a region growing procedure, in which thegrey-values of the pixels are used as a quantitativemeasure to control the shape of the regions grown.4) Now we consider the same segments, but in thecompensated image. We calculate the mean grey-value, and the “contrast” of each segment. The “con-trast” is a second-order statistic, and is frequentlyused in texture analysis. Its precise definition is givenbelow.5) Again we consider the same segments obtained instep 3), but now in the original image. Now we in-crease or decrease the grey-value of each pixel in suchsegment, depending on the grey-value of some of itsneighbouring pixels. The exact way the grey-values

5

are changed, as well as the number of times this pro-cess is performed is dependent on the contrast and themean grey-value of the segment as calculated in step4).

We shall describe and motivate each step in detail:1) Construction of the “Compensated Image”: Whenmaking an ultrasound image of the neonatal brain,the sonologist can select various scanner settings, likethe power (the amplitude of the emitted waves), thegain (overall amplification of the received signal), thedepth (the depth on which the emitted ultrasoundbundle is focussed), the Time Gain Compensation(different amplifications of the reflected signals fromdifferent depths) etc. Since we want the filter to dis-tinguish between various tissues, and want to use firsten second order statistical parameters for that (whichare obviously influenced by these scanner settings), wehave to construct a “standard image” first; an imagewhich is independent of those scanner settings. In [1]extensive study of this problem has been made, anda compensation algorithm which constructs such kindof standard image is described.2)Application of a Median Filter: Because we are go-ing to apply a region growing procedure based on thegrey-value of the pixels, we want to get rid of thesharpest speckles first without blurring the image toomuch. Otherwise the regions will “grow around” thesespeckles. To achieve this, we apply a median filterwith an 11x1 window to the compensated image. Thereason that we choose these sizes for the window, isthat, as a result of the limited lateral resolution ofthe imaging equipment, most of the speckles appearas short, thin, horizontal lines. So we should make thewidth of the window as small as possible. We foundan optimal length of 11 by experiment.3) Initial Segmentation: Texture information is bydefinition region based. Since we want to extract tex-ture information (contrast) from the image, our firststep is to make an initial segmentation. As the meangrey-level of is one of the distinctive parameters, weachieve this segmentation by a region growing proce-dure controlled by the grey-value of the pixels.

All pixels are checked from left to right, from topto bottom. The first pixel (i, j) is taken as the initialseed pixel. Its grey-value is denoted as α(i,j). Then itseight neighbouring pixels are considered. Wheneverthe grey-value of one of these pixels differs less thana fixed number ∆α from α(i,j), and the pixel has notbeen classified in a different region already, then itis added to the region under construction. Here ∆α

Fig. 11. A region

(the tolerance) is one of the adjustable parameters ofthe filter. In the results shown in this paper, we set∆α = 2.

After that, these new members of the region arechecked in the same way, the grey-values are still com-pared with α(i,j), etc. In a separate array, we keeptrack of which pixels already belong to a region.

So, after a region stopped growing, (because all sur-rounding pixels already belong to other regions or be-cause none of the adjacent pixels has a grey-value thatfalls within the accepted range of tolerance), the firstnext pixel which does not belong to one of the regionsalready formed, is taken as a new seed pixel, and thesame procedure is repeated.

In short, to test whether a pixel (m, n) belongs tothe homogeneous region of a seed pixel (i, j), the fol-lowing must be satisfied:

• Pixel (m, n) is “connected” to pixel (i, j)• |α(i,j) − α(m,n)| < ∆α• Pixel (m, n) does not belong to a region, which hasalready been formed before.

4) Calculation of the Various Parameters: For ev-ery region in the compensated image we calculate themean-grey value and the contrast. The contrast is de-fined as follows: suppose we have a region r, like, forinstance, the one shown in figure 11. We denote byAkl the number of pairs of adjacent pixels with grey-values k and l respectively. (In our example Akl = 3).Then the contrast γr of the region r is defined as:

γr =∑255

k,l=0(k − l)2Akl∑255

k,l=0 Akl

.

Practically, the contrast is calculated by means of thecooccurrence matrix. See, for example, [4] and [5] formore details about this.

5) Smoothing Operation: The actual smoothingprocedure is applied to the original image. Since wehave segmented and classified the regions now, we canfilter the various regions adaptively: in the healthy re-gions we suppress the light speckles. In the ill regions,

6

whose overall appearances are lighter, we create moreuniformity by removing the “black speckles”. The fil-ter itself is a modified version of the one presentedin [3]. Actually we work with two slightly differentprocedures. We will call them the “lightening pro-cedure” and the “darkening procedure”. Both workin three consecutive steps: North South adaptation(NS step), North-East South-West adaptation (NE-SW step), North-West South-East adaptation (NW-SE step).

NS step We work with two images; the second imageis an exact copy of the original image. Every pixel inthe original image is scanned. In the darkening pro-cedure, we first check if the grey-value of the pixelunder consideration is smaller than that of its north-ern neighbour or that of its southern neighbour. Ifthat is the case, then the grey-value g of the pixel atthe same position in the second image is increased by1. After that we check if the grey-value of the pixelunder consideration is higher than that of its north-ern or that of its southern neighbour. If so, then wedecrease the (current) grey-value g′ of the pixel at thesame position in the second image by 3.

In the lightening procedure, we work the other wayaround. First we check if the grey-value of the pixelunder consideration is higher than that of its northernneighbour or that of its southern neighbour. If thatis the case, then the grey-value g of the pixel at thesame position in the second image is decreased by 1.After that we check if the grey-value of the pixel un-der consideration is lower than that of its northern orthat of its southern neighbour. If so, then we increasethe (current) grey-value g′ of the pixel at the sameposition in the second image by 1.

Obviously, the darkening procedure has an inclina-tion to darken the pixels (it only whitens a pixel ifboth its neighbouring pixels are lighter, and it doesnot change its grey-value if all three pixels have ex-actly the same grey-value. In all other cases it dark-ens it). The lightening procedure has an inclinationto lighten the pixels, and acts a bit less strongly.

After having scanned all pixels of the region, thesecond image thus constructed is used as the inputimage for the next step.

NE-SW step, NW-SE step Analogous to the NSstep for both procedures, but now the grey-valueof the pixel under consideration is compared withrespectively the North-Eastern and South-Westernneighbour, and the North-Western and South-Easternneighbour.

As already mentioned the speckles in the image ap-

contrast

lighting procedure

darkening procedure

number

of

iterations

Γ ΓB T

0

1

2

3

-1

4

Fig. 12. Determining sr

pear as short, thin, horizontal lines, so no comparisonin the East-West direction is made.

When applying the filter, we can adjust four param-eters: B (bottomcontrast), T (topcontrast), M (max-imum number of iterations), and L (limitgrey-value).The visual interpretations of our measurements, pre-sented in figure 4, leads us to set these as follows:

B = 10, T = 40, M = 4, L = 67.

Let r be the region under consideration, µr its themean grey-value, γr its contrast. Now, we calculate avariable sr as follows:

• If µr > L or γr ≥ T , then sr = 0.• If µr ≤ L and γr ≤ B, then sr = M.• If µr ≤ L and B < γr < T , then

sr = �(T − γr) · MT − B

�.

So, if µr ≤ L, then sr is dependent on the contrastas shown in figure 12.

Finally, we apply the lightening or the darkeningprocedure on the region iteratively. Which of the twoprocedures is chosen and how many times it is applieddepends on the value of sr. If sr ≤ 1, then we applythe lightening procedure M times; if sn > 1, then weapply the darkening procedure sr times.

V. results

In figure 13 the result of the filter on the imageof figure 6 is shown. In figure 15 we show the final

7

Fig. 13. Filtered image

Fig. 14. Initial snake position

result of the snake from the same start position as inthe unprocessed image (figures 14 and 8). To simplifycomparison, we laid the final position of figure 15 overthe original image (figure 16).

VI. conclusion

Visual comparison of the results leads us to the con-clusion that the application of the GVF snake on theprocessed image performs considerably better with re-gard to finding the shape of the region of interest aswell as with regard to finding the correct area, thanon the unprocessed image. Furthermore, the proposedtechnique works fast and is computationally not inten-sive.

VII. References

[1] B. Simaeys, W. Philips, I. Lemahieu, P. Govaert,“Quantitative analysis of the neonatal brain by ultra-sound”, Computerized Medical Imaging and Graphics,

Fig. 15. Final snake position

Fig. 16. Copied on the original image

vol. 24 (2000), p.p. 11-18

[2] O. Basset, Z. Sun, J.L.Mestas, G.Gimenez, “Tex-ture analysis of ultrasonic images of the prostate bymeans of co-occurrence matrices”. Ultrason. Imag.,vol. 15, p.p. 218-237 (1993)

[3] T. Crimmins, “Geometric Filter for Speckle Re-duction”, Applied Optics, vol. 24, no. 10 (15 may1985)

[4] Yassar M. Kadah, Aly A. Farag, Jacek M. Zurada,Ahmed M. Badawi, Abou-Bakr M. Youssef, “Classi-fication Algorithms for Quantitative Tissue Charac-terization of Diffuse Liver Disease from UltrasoundImages”. IEEE Trans. on Med. Imaging, vol. 15, no.4,(august 1996), p.p. 466-478

[5] Y-N Sun, M-H Horng, X-Z Lin, J-Y Wang, ”Ul-trasonic Image analysis for Liver Diagnosis”, IEEE inMedicine and Biology, nov/dec 1996, p.p. 93-101

8

[6] Chengyang Zu, Jerry L. Prince, “Gradient VectorFlow: a new external force for snakes.”, Conf. onComp. Vis. Patt. Recog (CVPR’97), p.p. 66-71.

[7] Chengyang Zu, Jerry L. Prince, “Snakes, shapes,and Gradient Vector Flow”, Accepted in the IEEETrans. on Im. Proc.