Texture analysis of foot sole soft tissue images in diabetic neuropathy using wavelet transform M. Puri I K.M.

Patil I V. Balasubramanian 2 V.B. Narayanamurthy 3 1Biomedical Engineering Division, Department of Applied Mechanics, Indian Instit ute of Technology, Madras, India 2Department of Biotechnology, Indian Institute of Technology, Madras, India 3Diabetic Foot Clinic, Sundaram Medical Foundation, Anna Nagar, India Abstract--The paper presents a new method of characterisation of texture changes in foot sole soft tissue ultrasound (US) images, as observed to occur in diabetic s ubjects, using wavelet transforms. US images of the soft tissue subcutaneous layer were t aken with a 7.5 MHz linear transducer probe placed parallel to the skin surface. The foot sole hardness was characterised by Shore level. A 2D discrete wavelet transform was p erformed on the US images to extract features that encode the internal state of the foot sole soft tissue. The global energy feature computed at the output of each wavel et channel was found to achieve excellent delineation between the normal and the diabetic groups. An important finding was a strong correlation, in the order of 0.84 and above, between the feature values that reflect changes in the internal arrangeme nt of the tissue, and the externally measurable hardening of the skin, characterised b y the Shore levels, with the latter known to be high for diabetics. A comparison drawn between diabetic ulcer and non-ulcer groups established a change in the order of 122-311% in the textural parameter, as influenced by a corresponding 66.7-200% change in the respective Shore values. Thus US examination of foot sole soft tis sue and its texture analysis may serve as sources of valuable information regarding the internal changes taking place with progressive hardening of the soft tissue and thereby help the clinician in taking appropriate preventive measures. Keywords--Diabetic neuropathy, Foot sole hardness, Foot sole soft tissue ultraso nography, Wavelet-based texture analysis, Tissue classification Med. Biol. Eng. Comput., 2005, 43, 756-763 1 Introduction DIABETES CONTINUES to be one of the most common factors associated with lower-extremity amputation in postindustrialised and developing countries. Diabetic foot ulceration leads to life-or limb-threatening complications and is the single most common precursor to amputation. Diabetic foot ulcers have been identified as major factors in 85% of lowerextremity amputations (BIRKE et al., 1995). It has also been found that the geometric and material properties of the foot sole soft tissue change prior to plantar ulcers (ROBEI~TSON et al., 2002). Such changes in the properties of the foot sole soft tissue result in increased tissue stresses and stress gradients (THOMPSON et al., 1999; THOMAS et al., 2004). The foot sole hardness values (as characterised by Shore levels) increase in certain foot sole areas prone to ulcer development (CHARANYA et al., 2004a) and, hence, have been measured for each such area in our case. The skin-fat interface has been reported as a likely site of initial ulcer formation owing

Correspondence should be addressed to Emeritus Professor K. M. Patil; emaih patilkm@yahoo.com Paper received 6 April 2005 and in final form 22 August 2005 MBEC online number: 20054059 9 IFMBE: 2005 756 to increased peak first principal stresses at this location with changes in its material properties (THOMPSON et al., 1999). Therefore a reliable, non-invasive method of early detection of such structural changes occurring at the skin-fat interface is clearly desirable, for timely diagnosis of the possible existence of foot sole soft tissue ulcers. Diagnostic ultrasound has been a useful clinical tool for imaging organs and soft tissues in the human body for more than two decades (WELLS, 1977). This work attempts to perform differentiation between normal and abnormal foot sole soft tissue based upon the examination of B-scan images. However, for diagnosis of the progress of diabetes, particularly in its early phase, the main obstacle is the very subtle visual difference between sonograms of normal and abnormal foot sole soft tissue. One possible approach, as adopted in this study, can be the use of texture analysis to extract meaningful features or specific image properties and categorise or classify such sonograms. This is because structural changes, observed in advanced cases of diabetes, lead to changes in the acoustic properties of the foot sole soft tissue. Such changes produce differences in textural pattern, as detected by ultrasound, and can be observed between normal and diabetic subjects. In the field of computer vision, texture plays an important role in low-level image analysis and understanding. There is Medical & Biological Engineering & Computing 2005, Vol. 43 no formal or unique definition of texture, making texture analysis a difficult and challenging problem. The classification and segmentation of texture content in digital images have received considerable attention during the past decades, and numerous approaches have been presented. Statistical, model-based and signal processing techniques are the most commonly used approaches. The focus of this paper will be on multirate and multiresolution signal processing approaches. A common denominator for most signal processing approaches is that the textured image is submitted to a linear transform, filter or filter bank, followed by some energy measure. To accomplish multiresolution decomposition, a number of related techniques have been developed, including Gabor, Haar, Walsh-Hadamard expansions, Gaussian and Laplacian pyramids, sub-band filtering and many more. However in the last 15 years, wavelet methods have emerged to provide a more formal, solid and unified approach to multiresolution representations (MALLAT et al., 1989). Medical images, e.g. ultrasound, have been observed to convey useful diagnostic information through their visual texture. Some work carried out on texture analysis of ultrasound images is reported in REATH et al. (1985), WAGNER et al. (1986) and GARRA et al. (1989). An effective and separable extension of the 1D wavelet transform to the 2D case has been proposed and applied to ultrasound images of liver tissue for the detection of cirrhosis in its early stages (MoJSILOVIC et al., 1996). In this study, we have utilised the two-dimensional discrete wavelet transform to obtain appropriate features that quantify

texture changes in ultrasound images of diabetic foot sole soft tissue compared with those of normal subjects, which, in appearance, are inaccessible to unaided human appreciation. 2 Methodology To examine changes in tissue echogenicity due to the altered material properties of the foot sole in diabetes, US images were recorded, together with measurements taken on the foot sole to determine its hardness, for both normal and diabetic subjects. These US images were then passed through a filtering stage consisting of a dyadic discrete wavelet transform. Textural features were computed on these filtered images to provide a distinction between the normal and diabetic foot sole soft tissues. 2.1 Foot sole hardness measurement (Shore level) The instrument used for measuring the hardness of the foot sole is the Shore meter, or durometer or hardness tester (as per ASTM-D 2240 standards). It is used to quantify the levels of hardness of soft materials such as rubber or soft tissue. The Shore meter works on the principle of indentation using a truncated cone-tipped indenter. For each subject, normal and diabetic, measurements of the hardness of the foot sole were carried out in all foot sole areas prone to plantar ulcers (ARMSTRONG et al., 1998), namely, the medial heel (area 1), the lateral heel (area 2), the first metatarsal (area 5), the lateral metatarsals (area 7) and the big toe (area 8), as indicated in Fig. 1. The demographic details of the control and diabetic subjects are given in Table 1. 2.2 Foot sole soft tissue image acquisition For the purpose of the study, ultrasound images were recorded in B-mode by a 7.5 MHz linear array transducer being placed over the skin of the foot sole, parallel to the skin surface. US images were scanned, and the image of the tissue below the bone was captured for study. All US recordings were made on subjects in the standing position or loaded foot sole soft tissue conditions. As shown in Fig. 2, the transducer was oriented either from the lateral to medial or medial to Medical & Biological Engineering & Computing 2005, Vol. 43 8 9 10 5 6 7 4 3 1 2 Fig. 1 Foot sole areas lateral direction (depending upon the foot region scanned) for all five ulcer-prone foot sole areas mentioned above. The US images were recorded for normal as well as diabetic subjects. 2.3 Two-dimensional discrete wavelet transform definition Multiresolution techniques intend to transform an image (or any function, in general) into a presentation in which information regarding both the nature of the frequency components (high or low) and the location of occurrence of these frequencies in the image axe preserved. Therefore such techniques offer both frequency and time localisation, as opposed to only frequency-domain decomposition by Fourier transforms. The discrete wavelet transform (DWT) analyses a signal based on its content in different frequency ranges. Therefore it is very useful in analysing repetitive patterns such as texture. The wavelet transform is expressed as a decomposition of a signal f(x) ~ L2(R) (set of square integrable functions) into a family of functions that are translations and dilations

of a mother wavelet function ~(x). The function qrs(x), given by qrs(x)= ~/sqr(sx), where s ~ R 2, generates a dilated version, if s < 1, and a compressed version, if s > 1, of the mother wavelet function 0(x). Similarly, 0,(x)= ~(x- a) where, a ~ R 2, produces a right translation, if a > 0, and a left translation, if a < 0, of the mother wavelet function O(x). Therefore combining the above two definitions into one, ~0x(X) = ,./SO (s(x - a)) (s, a C R 2) generates a family of wavelets corresponding to translations and dilations of the mother wavelet ~(x). The continuous wavelet transform of a function f(x) is defined as W f ( s , a ) = I ' ~ f ( x ) ~ / s ~ t ( s ( x - a ) ) d x (1) As the continuous wavelet transform is redundant, it can be discretised by sampling the scale parameter s and translation Table 1 Demographic details of control and diabetic subjects Control Diabetic* Male:female 4 : 4 12 : 4 Number of feet 8 16 Average age, years 45.4 _+ 10.6 56.8 _+ 12.5 Duration of diabetes, years 10.6 -t- 7.7 Number of feet with ulcers 3 Ulcer duration, months 9.16 _+ 7.9 Sites of ulcers Area 2 (lateral heel) 1 Area 5 (I MTH) 1 Area 8 (big toe) 1 *All diabetic subjects were non-insulin dependent, all subjects withulcers were neuropathic Fig. 2 Ultrasound probe placed over foot sole soft tissue, parallel to skin surface (in this case, held along medial-to-lateral direction) parameter a. The most common choices axe s = 2 J (dyadic scales) and a = n/2 j, where j, n E Z j. Although a wavelet family can be built for sequences of scale other than (Z2)jcz2, dyadic scales lead to simpler decomposition algorithms (MALLET, 1989). Inserting these values into (1) yields the DWT of the signal f (x) as Wa(j, n) = (((x), Om(x - 2~n)) (2) where <...> denotes the inner product. An important class of wavelets qt(x) E L2(R) has been shown to exist, such that (4'2J(x - (n/2J)))(j, n c R2) constitutes an orthonormal basis of L2(R). These particular wavelets are called orthogonal wavelets, and the transform is called an orthogonal wavelet transform. Therefore any function now can be easily reconstructed from its decomposition into a wavelet orthonormal basis using the classical expansion formula of a vector into an orthonormal basis, given by f(x) = Z Z (f(x), 4,~(x - 2~n))g'~(x - 2~n) jEz nEz (3) The above equation can be easily extended into two dimensions for image processing applications (MALLAT et al., 1989). For multiresolution decomposition of images, it is often desirable to differentiate the local orientation of the image features. For this purpose, a scaling function qS(x, y), which is a lowpass filter, is introduced along with three wavelet functions i _~ .~ 1 2 ~O(x, y3) , (1 -.~ 1 -.~ 3), where each wavelet ~O (x, Y)O (x, y) and 4' (x, y) can be interpreted as the impulse response of a bandpass filter having a specific orientation selectivity in the vertical, horizontal and diagonal directions, respectively. The

low-pass and high-pass filtering actions axe performed using digital filters with impulse responses G and H, respectively. To achieve decomposition as shown in (2), it has been shown that G and H must necessarily form a pair of quadrature mirror filters (SMITH and BARNWELL, 1986). An imagef(x, y) with a spatial resolution of 2 j can therefore be decomposed through level-1 using f(x, y) = Al q)2J(x - 2Nn, y - 2Nm) + H 1 @ (x - 2Jn, y - 2Nm) + Vl~22j(x- 2Jn, y - 2Nm) + Dl@22j(x - 2Jn, y - 2Nm) (4) 758 where A 1 = ( ( f ( x , y), d)v(x - 2~n, y - 2~m)))n, m~z 2 (5) H 1 = ( ( f ( x , y), O}(x - 2~n, y - 2~m)))n, m~z 2 (6) V 1 = ( ( f ( x , y), @(x - 2~n, y - 2~m)))m ,~z 2 (7) D 1 = ( ( f ( x , y), @(x - 2~n, y - 2~m)))m ,~z 2 (8) Each of the above sequences of inner products can be considered as an image. The DWT decomposition of an image into four channels, namely A 1, H 1, V 1 and D 1, involves first the convolution of the original image with impulse responses of low-pass filter G and high-pass filter H, respectively, along the rows and then columns, which produces four filtered images, each with every second sample redundant. Therefore the filter stage is followed by the process of sub-sampling by a factor of 2 (discarding every other sample), which reduces the size or spatial resolution of the filtered images to half the original image. Therefore level-1 DWT decomposition of an image produces a representation of the image in the form of four sub-images, where A 1 represents the spatial distribution of low-frequency components, and H 1, V 1 and D 1 represent the spatial distribution of high-frequency components present at a resolution half that of the original image. Among the high-frequency channels, H 1 gives the horizontal edges (vertical high frequencies), V 1 gives the vertical edges (horizontal high frequencies), and D 1 gives higher frequencies in both directions (corners). This set of four images is called an orthogonal wavelet representation in two dimensions. The decomposition process can be recursively applied to the low-frequency channel A 1 to generate image details A 2 (low-frequency channel) and H 2, V 2 and D 2 (high-frequency channels), at the next level. Such a method of decomposition, where images at one level are at a resolution equal to half that of the previous level is called a pyramid structured wavelet decomposition (PSWD). The G and H filters axe used to implement the wavelet transform. Although a wide range of wavelet basis functions have been discovered to date, all DWT decompositions in this study were performed using the commonly used Daubechies eight-tap filter (DAUBECHIES, 1992). CHANG and KUO (1993), in their studies, recommend the use of dominant frequency channels with large energy values for the purpose of texture discrimination. Thus, for our study, we performed feature extraction from the level-1 wavelet-based frequency channels, because greater texture energy is found to exist in these channels. 2.4 Application of DWT for texture analysis of foot sole US images Ultrasound images of the foot sole soft tissue used in this research were all captured from the same machine and then

digitised with 480 576 pixels and 256 grey-level resolution. As explained in Section 2.2, ultrasound images of five foot-sole areas (namely areas 1, 2, 5, 7 and 8) for the feet of normal and diabetic subjects were scanned. The sets of data for normal subjects consisted of 40 feet areas from eight feet, and, for diabetic subjects, they consisted of 80 feet areas from 16 feet. As this study attempts to address the discrimination ability of image texture to detect early changes occurring at the skin-fat interface, we considered diabetic patients in different stages of the disease. From each image, a region of interest (ROI) of size 64 64 pixels (approximately 1 cm 1 cm in actual dimensions) was selected. ROIs were chosen to include only the foot sole soft tissue, without blood vessels, acoustic shadowing or any type of distortion. Because the ultrasound images were taken by different clinicians, to eliminate the effects of unequal gain settings, the mean of each image was removed, before processing. Upon the discrete wavelet transform being applied to each of the ROI images, four sub-images, each equal to half the spatial resolution of that of the original, were obtained at the output. The energy responses used as texture measures were then calculated for all the four DWT filtered images, namely A 1, H 1, V 1 and D 1, using ~/2 m/2 E-A1 = Z Z (AI(i' j))2 (9) i=1 j=l n/2 m/2 E-H1 = Z Z (Hl(i' j))2 (10) i=1 j=l n/2 m/2 E_V 1 = ZZ(V](i, j))2 (ll) i=1 j=l n/2 m/2 E-D1 = Z Z (DI(i' j))2 (12) i=1 j=l where, n and m are equal to 64 (size of each ROI image), and E_A 1 , E_H,1 E_V 1 and E_D 1 each represent energies in the A 1 , H 1, V 1 and D 1 images, respectively. frequency channels, namely A 1, V 1, H 1 and D 1, are represented in the form of an energy map in Figs 4c and f. As can be seen from the energy maps presented, there are significant differences between the normal and diabetic tissue, as shown by a significant numerical variance among their respective frequency channels. Figs 5a and b, respectively, show the spatial intensity distribution in area 5 for a normal subject and a diabetic subject having an ulcer at this location. The plots clearly indicate high spatial frequency power in the US images of the ulcer regions of diabetic subjects, compared with normal subjects. Fiigs 5c and d represent the corresponding distributions in the A energy channel after level-1 decomposition of these images. The high energy response of the A 1 sub-image, as shown in Fig. 5d, compared with Fig. 5c, indicates larger low-frequency variations across US images of diabetic subjects than those of normal subjects. Table 2 represents the energy concentrations in DWT filtered images corresponding to approximation (A 1) coefficients present in the US tissue images of normal and diabetic feet. Using statistical tests, comparisons of texture features axe made between normal and diabetic subjects with different

levels of foot sole hardness or Shore values. Statistical analysis using a Welch ANOVA post hoc test followed by the Dunnett t-test for parametric energy responses, calculated at A 1 , H,1 V 1 3 Results An image and its level-1 dyadic decomposition axe shown in Figs 3a and b, respectively. The LP and HPfilters axe used to implement the wavelet transform. This results in an output with the same size as that of the input. Fig. 3d shows a level-1 DWT decomposition of a sample image in Fig. 3c. As shown in Fig. 3b, A 1, H 1, V 1 and D 1 refer to level-lwavelet decomposition coefficients respectively, of any image. Figs 4a and d illustrate the US images of foot sole soft tissue for a normal foot and a diabetic foot, respectively. The level-1 DWT decomposition of these images is shown in Figs 4b and e. Energy calculations made on outputs of level-1 wavelet

9 . ~. -_~.: _..-~...=: : g.~-.:-.~-: : . 7 - ~ _ ' ~ - - - , ' - . "

9 .. -

9 - . , . , ,a- - ~_1, ~.. '.r.. "i"T,'" "'>" " =

9 . * .q b image . . , , , . m A 1 H 1 V 1 D 1 --- ...,=~.-, r .- :,~', . .

9 ~ . . ~ : . 1-'~ :

9 ~.%-.:~ " .= 504.1 1.1 10 4 d Fig. 3 (a) Input image; (b) decomposition at level-l; (c) sample texture image; (d) DWT level-1 coefficients using Daubechies 8-tap filter Medical & Biological Engineering & Computing 2005, Vol. 43 Fig. 4 1596 113 23 11.5 (a) Normal foot sole soft tissue US image in area 5; (b) level-1 DWT decomposition of(a); (c) energy map of(b); (d) diabetic foot sole soft tissue US image in area 5; (e) level-1 DWT decomposition of (d); (f) energy map of (e) 759