Skin Research and Technology 2005; 11: 79–90 Printed in Denmark. All rights reserved
Copyright & Blackwell Munksgaard 2005
Skin Research and Technology
Automatic measurement of dermal thickness from B-scan ultrasound images using active contours Jean-Michel Lagarde, Je´ro´me George, Romain Soulcie´ and David Black Centre Jean-Louis Alibert, Institut de Recherche Pierre Fabre, Toulouse, France
Background/purpose: Measurement of dermal thickness is useful in the evaluation of dermo-cosmetics for assessing not only morphological changes but also mechanical properties of this layer. Our aim was first to standardise the manual dermal thickness measurement procedure on B-scan ultrasound images, then to develop an automatic operator independent method to detect the boundaries of the dermis. Material and methods: The Dermcup s 20 MHz B-scan ultrasound system was used. The method used for detecting the boundaries was adapted from active contour algorithms. The innovative aspect of the method consists in an automatic initialisation of the first step of the algorithm. To validate the method, we correlated measurements obtained
by the manual and automatic approaches from a set of images from different anatomical sites. Results and conclusion: The results showed for the two measurement methods, 72% of the images were perfectly correlated. The remaining images required manual initialisation of the boundaries by a non-expert operator before the active contour process could be used. Subsequent to this semi-automatic procedure, the correlation was very high.
of dermo-cosmetics, it is essential to be able to characterise the skin if it is to be better understood, and if progress is to be made in developing more efficient care products. In this aim, the visualisation and the measurement of certain parameters is necessary. As in other fields of medicine (cardiology, obstetrics, etc.) researchers have resorted to ultrasound imaging, which provides high resolution over the few millimetres thickness of the dermis while being non-invasive and inexpensive. The major problem encountered in the ultrasound characterisation of the dermis is the high variability of the results obtained. The determination of characteristic measurements such as the mean thickness of the dermis and its standard deviation must provide results that can be used in studies involving for instance, mapping of the dermis and its characteristics depending on the location on the body, the sex, the age, and so on. And also in a patient to follow the evolution of diseased dermis or of a treatment for which the effects are to be quantified, as is the case in the Centre Jean-Louis Alibert (CJLA).
The study of slight variations of the dermis therefore requires accurate measurements which, in addition, must not be subject to strong variability. But, the current measurement techniques, detailed in the following pages, suffer from two types of variability:
I
N THE FIELD
Key words: clinical assessment – skin – image processing – active contours – ultrasound imaging
& Blackwell Munksgaard, 2005 Accepted for publication: 15 July 2004
inter-operator variability: depending on the operator performing the measurements, the results will not be identical; intra-operator variability: even a single operator performing the measurements on several occasions will not find the same result. This can be put down to the fact that the human operator has a wide range of interpretation within which to locate the ‘measurement’. (Variability because of the measurement instruments themselves is considered to be almost unavoidable in in vivo applications and is out of the scope of the present study.) Decreasing the human interpretation errors to reach better repeatability involves the development of rigorous and precise protocols. Following them scrupulously implies maximising automation of the various steps, the error then being reduced to
79
Lagarde et al.
the machine’s interpretation. The technique currently used at the CJLA follows a relatively precise protocol. Firstly, the DERMCUP (ATYS, Lyon, France) measurements (This device was described in (1)) must be carried out only after checking: that air bubbles in the gel have been reduced to a minimum – bubbles generate a strong echo and, owing to their absorption of energy, can lead to the appearance of columns of shade beneath them, that the probe is held vertically so that it does not measure an oblique section of dermis, that the probe is not placed over a tendon or a vessel (avoiding the appearance of artefacts in the image that would disturb the measurement), Note: Although none of these points appear excessively difficult to respect taken one by one, in practice, it is quite challenging to respect them all simultaneously. Measuring the thickness of the dermis then consists of measuring the dermis at three points of the image with a vertical line (Fig. 1). Averaging the three thicknesses gives the thickness of that image. The increasing use of ultrasound in medical imaging has led to abundant articles on the subject of their segmentation. The natural noise (i.e. the noise naturally present in the tissues being scanned) and artificial noise (i.e. the noise produced by the measurement instruments themselves) generally present on ultrasound images requires the use of powerful segmentation techniques. Among them, we can mention, segmentation by texture analysis (2, 3),
related techniques based on co-occurrence matrices (4, 5), and measurement of homogeneity and contrast (6). Active contour segmentation techniques are also popular. The basis of this technique, reported in reference (7), has been reconsidered to propose new models. Reference (8) proposes multiresolution to deal with noise in cardiology ultrasound images, while article (9) defines the so-called ‘inflatable snake’ model to detect closed contours that will allow, from an estimation of the area to be segmented, parasitic contours to be avoided. In the present study, we have chosen to work with the ‘greedy algorithm’ defined in (10). However, in the particular field of ultrasound images of the dermis, the literature available is much poorer (the techniques proposed in the articles mentioned above have almost all been validated in cardiology or obstetrics); in spite of the quality of the images obtained in dermocosmetology (little artificial background noise), the scale of the images means that the techniques of smoothing and multi-resolution do not efficiently reduce the natural noise of the tissues. The techniques reported here are: ‘Mode A’ technique: A rapid technique involving direct estimation of the thickness of the dermis from an image of mode A type. Interpolation technique: This technique, described in (11) aims to extract the junction between the dermis and the hypodermis (i.e. the lower limit of the dermis) on series of images. It consists of manually tracing out the border on sections made regularly across the series, then interpolating the areas remaining between them.
Materials and Methods In this section, we present the initial active contours technique, as reported in the original article of 1988 (7) and the modifications that have been made to adapt it to DERMCUP images and enhance the detection efficiency. We then present a validation study of the dermis recognition results.
Fig. 1. Technique of dermis thickness measurement on an image, as used at the Centre Jean-Louis Alibert.
80
Principal of active contours: the classic model The technique of contour detection using active contours consists of marking out a curve, near the contour to be defined, which is then fitted, by an iterative convergence process to make this ‘artificial’ contour lie exactly over the real contour. The way in which the active contour moves
Measurement of dermal thickness from B-scan ultrasound images
during the process has led to it being called the snake. Behind this apparently simple principle is the following theory.
of the image: if the snake is designed to find high energy when placed in homogeneous areas of the image, it will necessarily seek, during minimisation, a position over the contours. If the points of the contours are classically associated to the maxima of a gradient, we will have, for example:
The underlying mathematics Mathematically, an active contour is modelled by a parametered curve v: v:
½0; 1 ! <2
PðvðsÞÞ ¼ kHðGt IðvðsÞÞÞk2
ð1Þ
s7!vðsÞ ¼ ðxðsÞ; yÞðsÞÞ
Smoothing by the Gaussian filter Gt softens the variations of potential and enables the contours to attract the snake from further into the image, but this is not ‘compulsory’. A third energy term can be added to take into account constraints such as an obligation to pass by certain points or to avoid other points. This term enables the user to guide the snake towards certain areas or to keep it away from erroneous contours. The link between the total energy function and the displacement of the snake during the convergence process is ensured by the Euler–Lagrange equations, which establish a relationship between the energies and the forces applied at each point of the curve. The snake’s energy function reaches a minimum when the resultant of the forces is nil. The corresponding equation is:
whose spatial derivatives are written: dv d2 v and vss ¼ 2 ds ds For any contour v, there exists an associated energy E(v) which the convergence process attempts to minimise. This energy can be decomposed into two terms: The internal constraints are a function with the form: Z1 h i aðsÞkvs ðsÞk2 þbðsÞkvss ðsÞk2 ds ð2Þ Eint ðvÞ ¼ vs ¼
0
The first term under the integral increases with the length of the snake, thus, during minimisation of the energy, this length tends to decrease. Similarly, the second term tends to limit the curvature of the snake because it is second order. Coefficients a and b, which weight the two terms, enable the elasticity and the suppleness of the snake to be parametrised. These snake constraints are called ‘internal’ because they are totally independent of the position of the snake in the image. The external constraints make the snake adapt to the image: Z1 ð3Þ Eext ðvÞ ¼ l ½PðvðsÞÞ ds
gvt
daðsÞvs d2 bðsÞvss þ ¼ HP ds ds2
VðtÞ ¼ ðA þ gIÞ1 :ðVðt 1Þ HPðVðt 1ÞÞÞ ð6Þ where V(t) is the vector of the coordinates of the points of the snake and A is a pentadiagonal matrix dependent on the values of a, b and g. For closed snakes with non-fixed extremities, A is written:
P is a scalar potential which corresponds, for instance, to a measurement of the homogeneity b
c
6b 6 6 6c 6 6 6 A ¼6 6 6 6 6 6 4c
a
b
c
b
a
b
:
c
b
:
:
:
:
:
:
b
c
:
b
a
b
c
b
a
c7 7 7 7 7 a ¼ 2a þ 6b 7 7 7 with b ¼ a 4b 7 7 c¼b c7 7 7 b5
c
b
a
c
c
b
3
a
b
ð5Þ
here t is the time variable and vt ¼ dv=dt the time derivative. Estimating the derivatives by finite differences, discretisation of the problem leads to the following matrix equation:
0
2
ð4Þ
81
Lagarde et al.
Note that this matrix is not reversible because (a1b1c) 5 0. Also, for the inversion of ðA þ gIÞ, g must not be nil. This matrix expresses the fact that each point of the snake will move during an iteration taking into account its former position (this corresponds to the ‘a’ diagonal), those of its immediate neighbours (coefficients ‘b’) and those of its neighbours’ neighbours (coefficients ‘c’). So, the triplets isolated from coefficients that appear at the top right and bottom left in matrix A indicate the influence of the last two points of the snake on the first two points, and vice versa, hence the contour obtained using this matrix is 2
where (xi(t), yi(t)) are the coordinates of the ith snaxel at instant t. As seen above, the convergence process consists of iterating equation (6) which moves the snake from its position at instant t to a new position at instant (t 1 gt); the process is repeated until the snake no longer moves. Each iteration occurs as follows: 1. Calculation of the matrix HP(V(t–1)) it is the vector of the outer force field along the snake: where F is the potential gradient, Fx its x component and Fy its y component, so we have:
Fxðx0 ðt 1Þ; y0 ðt 1ÞÞ
6 Fxðx1 ðt 1Þ; y1 ðt 1ÞÞ 6 6 6 HPðVðt 1ÞÞ ¼ 6 6 6 6 4
Fyðx0 ðt 1Þ; y0 ðt 1ÞÞ
3
Fyðx1 ðt 1Þ; y1 ðt 1ÞÞ 7 7 7 7 7 7 7 7 5
Fxðxn ðt 1Þ; yn ðt 1ÞÞ Fyðxn ðt 1Þ; yn ðt 1ÞÞ closed (we will see below how A has been adapted to measure the thickness of the dermis). The process of snake convergence therefore consists in iterating this last equation until the snake reaches a stable position. Calculus techniques The snake is discretized by points, called snaxels (from ‘snake’ and ‘pixel’) placed regularly along 2
aþg 6 b 6 6 c 6 6 A¼6 6 6 6 6 4 c b
b aþg b c
c
c b aþg b :
c c b : : : : : : b c
: b aþg b c
its contour. The coordinates of the snaxels at instant t are stored in a two-column vector V(t): 3 2 x1 ðtÞ y1 ðtÞ 6 x2 ðtÞ y2 ðtÞ 7 7 6 7 6 7 VðtÞ ¼ 6 7 6 7 6 5 4 xn ðtÞ yn ðtÞ
82
This vector is noted Fext(t–1). 2. Calculation of the matrix V(t 1) Fext(t 1) This matrix can be considered as a pseudosnake in which each of the snaxels is moved according to the force field of the image, but without taking the neighbours into account. It is noted V 0 (t 1). 3. Inversion of matrix (A1gI) This matrix is fairly empty because it is pentadiagonal: b c
c b aþg b
c b aþg
3 7 7 7 7 7 7 with 7 7 7 7 5
a ¼ 2a þ 6b b ¼ a 4b c¼b
Matrix inversion is an operation that is very costly in calculation time if a ‘general’ matrix inversion algorithm is used. However, invertible pentadiagonal matrices can ‘benefit from’ inversion with linear complexity using Choleski’s algorithm. This method, named after the mathematician who developed it, consists in decomposing the matrix to be inverted into the product of two
Measurement of dermal thickness from B-scan ultrasound images
matrices:
remain nil such that, during the convergence process, it is only free to slide up and down the first column like on a rail. Likewise, the last snaxel is blocked in the last column. This eliminates the problem of behaviour at the ends.
A¼L:U where L is a lower-triangular matrix and U an upper-triangular matrix (this decomposition is possible for invertible pentadiagonal matrices). We then have: A1 ¼U 1 : L1
Orientation of the force field of the image The role of the outer force field is to guide the snaxels towards the contours: it indicates, for all points on the image, ‘in which direction the closest contour is found’. The problem that is raised here is that on generating the force field from the classic potential
which accelerates calculations, triangular matrices being linearly invertible. It is important to note here that this matrix does not have to be calculated at each step of the convergence process: if the number of snaxels and the values of coefficients a, b and g are unchanged, this calculation can be avoided. 4. Final calculation of V (t) 5 (A1gI) 1. V 0 (t – 1)
Adaptation to
DERMCUP
PðvðsÞÞ ¼ kHðGt IðvðsÞÞÞk2 important information is lost: the orientation of the contours, i.e. even though the classic force field effectively guides the snake towards the contours, it does not allow a distinction to be made between contours of increasing intensity and contours of decreasing intensity. In other words, an image and its negative have the same force field. This situation is most unfortunate because the dermis appears brighter than the other elements of the image. It would therefore be better for the snake to seek a separation between ‘a bright area and a dark area’, avoiding the type of result illustrated in Fig. 2a.
images
Detection of non-closed contours
Adaptation of the classic model first requires a different approach for the ends because the classic model involves closed contours. The modification must be made on the pentadiagonal matrix A. As mentioned above, closure of the active contour is because of the presence in A of the two triplets of isolated coefficients: A is modified as follows: 2
a
aþbþc
6b a 6 6 6c b 6 6 c 6 A ¼6 6 6 6 6 6 4c
b
c
a
b
:
b
:
:
:
:
:
:
b
c
:
b
a
b
c
b
a
6 bþc c7 6 7 6 7 6 7 c 6 7 6 7 6 7 7!A¼6 6 7 6 7 6 7 c7 6 6 7 4 0 b5
c
b
a
c
b
2
c
b
c
3
b
In this way, the following phenomenon is simulated: at each end of the snake two ‘virtual’ snaxels are added to the last snaxel to act as neighbours for the two first and the two last snaxels that were ‘lacking’ neighbours. Alone, this simple adaptation would tend to disturb the extremities of the snake. Fortunately, the dermis contour crosses the whole width of the image, so we block the first snaxel in the first column of the image: its abscissa is made to
0
b
c
a
b
c
b
a
b
:
c
b
:
:
:
::
:
b
c
:
b
a
b
c
b
a
c
b
0
0
0
3
7 7 7 7 7 7 7 7 7 7 7 c 7 7 bþc 5 0
aþbþc
A suitable system was developed using the following function as force field: F:
<2 ðx; yÞ
<2 2 3 d Iðx;yÞ Fx ðx; yÞ ¼ d2 x 4 5 7 ! 2 Fx ðx; yÞ ¼ d dIðx;yÞ 2 y ! 2
This force field preserves the orientation of the contours (the principle is illustrated in one dimension in Fig. 3 which shows the phenomenon
83
Lagarde et al. a
b
contours
contours
image 1D
image 1D
derivation+valeur absolue
derivation
derivation
Result of the attraction force
Fig. 3. Principle of field orientation: (a) classic force field, (b) oriented force field.
orientation of the contours as seen by the snake must also be considered. If a snake is to work in an oriented force field, it too must be oriented, i.e. it must know how to locally distinguish its left from its right in order to be able to inverse the force field or one of its components if necessary. This local orientation is calculated for each snaxel, consulting the positions of its two immediate neighbours: the consultation operates via the construction of vector HP(V(t–1)). Initialisation mechanisms
Fig. 2. ‘Errors’ to be avoided during contour orientation: (a) error, (b) vertical orientation, (c) horizontal orientation.
clearly). Not losing the absolute value changes the contours that correspond to drops in light intensity that become ‘anticontours’ and will actually repel the snake. Figure 2 b–c presents the structure of such a force field for a 2D image: the attractive or repulsive character of a contour in fact depends on its orientation in the image – it is clear that the
84
The active contours process, it has been seen, is a powerful and accurate contour detection tool. Its main failing, however, in addition to complex setting up and fine tuning, is that to be efficient it requires sufficiently accurate initialisation – which is too often left in the hands of a human operator. Automating initialisation using other, lower level, segmentation techniques reduces the workload of the human operator and at the same time decreases the variability of the results obtained. Mechanisms to do this have been developed: they propose an initialisation of the contours to be detected (upper and lower limits of the dermis) while allowing the human operator to modify it via the interface if it is considered to be erroneous. Once this preliminary phase has been carried out, the convergence process can be started and leads to a stable position where the snakes must correspond to the edges of the dermis. Construction of the ‘profile’ of a
DERMCUP
image
The images studied are strongly structured: we know that the area to be isolated crosses from one side of the image to the other, with relatively
Measurement of dermal thickness from B-scan ultrasound images
horizontal limits. This provides first an idea that considerably reduces the issue of initialisation and which has proved to be quite acceptable for the majority of the images processed (below, we will see some specific cases and the adaptations they required). The active contours that will seek out these limits can reasonably be initialised as segments of a straight line crossing the image more or less horizontally. In addition, the strong echo from the dermis, which is because of its ability to bounce the ultrasound waves back, means that the area to be isolated stands out brightly in the images in contrast to the rest of the tissue that does not give a strong echo. This prompted a second idea: summarising the image to what we can call its ‘profile’. We call the profile of the image the function that attributes a mean intensity to each line of the image (see Fig. 4 a, b). profile : N ! < 0 x < height 7! grey level along line x It is from this profile only that initialisation of the upper and lower limits of the dermis will be determined. Note that the curve obtained resembles that of ultrasound scans in mode A that the DERMCUP can also provide and from which estimations of the dermis can be made. Although the human brain can easily establish the correspondences existing between the various areas of the image and the variation of the profile,
in its as-defined rough form, the profile is too noisy to be directly interpretable by a simple algorithm in spite of the fact that each of its values arises from the average value of a whole line of the image. After smoothing with a 1D Gaussian filter of suitable power, the curve becomes much more clearly readable (see Fig. 4c). It keeps all its essential characteristics (intensity peaks and troughs) and the interval corresponding to the dermis can almost be isolated after a single thresholding operation. Upper limit of the dermis
On a smoothed profile of an image, the upper boundary of the dermis appears as a sharp rise in intensity that follows the low-intensity area corresponding to the gel. It can also be underlined at this point that the upper limit corresponds, in the vast majority of cases, to the maximum of the profile’s derivative (see Fig. 4d). Lower boundary of the dermis
The lower boundary of the dermis is both more irregular than the upper boundary and more difficult to accurately locate on the profile. To do so, we will use the upper boundary that is simpler to identify. Thus, the initiation algorithm for the lower limit starts by identifying the upper boundary. It then shifts the image such that the upper boundary is tight up against the top of the image, then it calculates the profile and its derivative for this new image. Risk of erroneous results
The initialisation phase, can, in certain cases, prove to be erroneous, mainly when: the probe membrane is stuck to the surface of the skin (Fig. 5a), the dermis–hypodermis junction is indistinct or inexistent (Fig. 5b). A manual initiation phase is then necessary before being able to initiate the iterations of the active contours. Measurement then becomes semi-automatic. Extraction of the parameters Fig. 4. An image, its profile and its derivative: (a) the image, (b) the profile, (c) the smoothed profile, (d) the derivative.
As the main aim is to produce a tool able to measure the thickness of the dermis automatically, this function has been included in the
85
Lagarde et al.
software. Once the gel–dermis and dermis–hypodermis junctions have been detected, the average thickness of the image can be measured. To do so, the area delimited by the two curves is calculated.
Results and Discussion Repeatability with the old method To study intra-operator variability, we carried out the following experiment. The dermis thickness was evaluated over 5 days by a person trained to perform the technique, using 45 randomly classified images coming from foreheads, thighs and forearms in equal proportions. The results are illustrated in Fig. 6. Variations can be seen for various images, which can reach up to 23% of the average thickness. Account must therefore be taken of these variations, which are because of the subjectivity of manual measurement, when finally interpreting the correlation with automatic measurements. Validation To validate the image processing algorithms, we compared the manual measurements, made by our ultrasound specialists, with measurements provided by the software. Six hundred and twelve images obtained with the DERMCUP were analysed.
Fig. 5. Examples of images in which the initialisation algorithm of the upper border cannot operate correctly.
1. First, we used the software in the totally automatic mode, 2. then an inexperienced user corrected the images in which he estimated that the contours were erroneous. A comparison of the
3.0 2.5
mm
2.0 1.5 1.0 0.5 0.0 1 Fig. 6. Intra-operator variability.
86
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 45
Measurement of dermal thickness from B-scan ultrasound images 3.8
a
y = 0.8792x + 0.3096 R2 = 0.4652
3.3
mm
2.8
2.3
2.3
1.8
1.8
1.3
1.3
0.8 0.8
y = 0.9444x + 0.1528 R2 = 0.6097
3.3
2.8 mm
3.8
1.0
1.2
1.4
1.6
1.8 mm
2.0
2.2
2.4
2.6
0.8 0.8
2.8
Fig. 7. Correlation between manual and automatic measurements for all sites pooled.
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
mm
b
3.8 y =0.5088x + 0.8925 R2 = 0.0678
3.3
mm
2.8
different results obtained is illustrated in the graphs reported in Figs 7–10. In these graphs, we show the tendency curve as well as the correlation coefficient R2.
2.3 1.8 1.3 0.8 0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
mm
c
3.8
y = 1.0605x + 0.119 R2 = 0.2416
3.3
mm
2.8 2.3 1.8 1.3 0.8 0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
mm
d
3.8
y = 0.3053x + 1.5906 R2 = 0.0086
3.3 2.8 mm
The automatic mode illustrated in Fig. 7 shows the results to be encouraging but not completely satisfactory because the cloud of points is strongly dispersed around the theoretical perfect straight line of the equation y 5 x. Separating the correlations into groups based on the location where the image was taken (Fig. 8), it is quickly seen that the only area where the figures are significant is the forearm. On the other hand, in the semi-automatic mode (Fig. 9), the slope of the straight line is very close to 1 and the gap at the origin is small, indicating that the values are accurate. The correlation coefficient is also close to 1 so the values are very closely correlated. Similarly, these excellent results remain when the images are grouped according to the site where they were taken (Fig. 10). The values generated by the program therefore seem to be justified and can be used, although the entirely automatic mode will not be sufficient in all cases. During this study, 111 images had to be corrected (19 for the forearm, 40 for the thigh, 23 for the forehead and 29 for the neck), i.e. 18% of the total number of images studied. The correction applied consisted of reinitialising the plot of a coherent boundary point by point before starting the algorithm again. For 13 of the images, the poor results were directly related to the low quality of the image (insufficient gel between the probe and the epidermis) and the images were discarded. The others arose from artefacts
2.3 1.8 1.3 0.8 0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
mm
Fig. 8. Correlation between manual and automatic measurements per site: (a) arm, (b) thigh, (c) forehead, (d) neck.
present in the images (very ultrasound-dense regions) that induce divergence of the algorithm (Fig. 11).
87
Lagarde et al. 3.8
y = 1.0347x + 0.0473 R2 = 0.9406
3.3
a
2.8 mm
2.3
2.3
1.8
1.8
1.3
1.3
0.8 0.8
y = 1.001x + 0.0793 R2 = 0.7853
3.3
2.8 mm
3.8
1.0
1.2
1.4
1.6
1.8 2.0 mm
2.2
2.4
2.6
0.8 0.8
2.8 b
Fig. 9. Correlation between manual and semi-automatic measurements for all sites pooled.
1.0
1.2
1.4
1.6
1.8 mm
2.0
3.8
2.2
2.4
2.6
2.8
y = 0.9399x + 0.2486 R2 = 0.7139
3.3
Influence of the method The graph shows that the mean deviation between the manual method and the automatic method is generally quite low. It can also be seen that the values of the automatic measurements are slightly higher than the manual results. To try to understand the phenomenon, we studied the method of the local expert and our conclusions are illustrated in Fig. 14. The manual method consists of measuring the gap between the two peaks present in the images on either side of the contour. The automatic method works to stabilise the active contours on low values of the second derivative of the
88
2.3 1.8 1.3 0.8 0.8
c
1.0
1.2
1.4
1.6
1.8 mm
2.0
3.8
2.2
2.4
2.6
2.8
y = 1.0049x + 0.0903 R2 = 0.6575
3.3
mm
2.8 2.3 1.8 1.3 0.8 0.8
d
1.0
1.2
1.4
1.6
1.8 mm
2.0
3.8
2.2
2.4
2.6
2.8
y = 1.067x + 0.0093 R2 = 0.8063
3.3 2.8 mm
Influence of the sites The parts of the body where the images were taken could have been without effect on automatic extraction, but owing to their dissimilar characteristics, they led to large differences (as seen in the above results) affecting the way the software behaves. The arm is generally dense and highly contrasted compared with the background, and its contours are quite clear. The lower limit of the thigh dermis, however, is much less clear and lacks ‘linearity’. The forehead is somewhere between the two but is particular in that it overlies the skull (Fig. 12). Figure 13 illustrates the difference in thickness between the sites. The mean thickness of the dermis on the forearm is about 1 mm, while on the thigh it is almost 2 mm. The differences in the values of the standard deviations in Fig. 13 demonstrate the difference in variability between people for the various sites. The forearm appears to be more uniform than the three other sites. All these results can be found in the groups of data representing the two methods of measurement.
mm
2.8
2.3 1.8 1.3 0.8 0.8
1.0
1.2
1.4
1.6
1.8 mm
2.0
2.2
2.4
2.6
2.8
Fig. 10. Correlation between manual and semi-automatic measurements per site: (a) arm, (b) thigh, (c) forehead, (d) neck.
signal and therefore homes in on the ‘centre’ of the contour. This difference in approach is probably sufficient to justify the more or less constant discrepancy between the manual measurements and those of the automatic system.
Measurement of dermal thickness from B-scan ultrasound images
Fig. 12. The different sites: (a) arm, (b) thigh, (c) forehead, (d) neck.
2.5
mm
2.0 1.5 1.0 0.5 0.0
a
b
c
d
e
Fig. 13. Influence of the site on the measurement: (a) arm, (b) thigh, (c) forehead, (d) neck, (e) all sites pooled.
Fig. 11. (a) Erroneous image, (b) point by point correction, (c) new detection.
Conclusion Bearing in mind the intra-operator variability analysed in the first part of the validation, it can be concluded that the correlation between automatic and manual measurement is excellent.
Fig. 14. Explanation for the discrepancy between automatic and manual measurements.
89
Lagarde et al.
Processing time, which is very long in the studies carried out at our centre, is greatly reduced. The time required by our specialist is 5–10 min per image while it is 1.5 s for a PC working at 1.8 GHz in the automatic mode. Correction time was, on average 30 s for about 18% of the images. The results presented here demonstrate that the method of automatic measurement of dermis thickness on 20 MHz ultrasound images, based on an active contours algorithm, can totally replace the manual method.
References 1. Diridollou S, Pavon-LeTraon A, Maillet A, Bellossi F, Black D, Patat F, Lagarde JM, Berson M, Gall Y. Characterisation of gravity-induced facial skin oedema using biophysical measurement techniques. Skin Res Technol 2000; 6: 118–127. 2. Kim N, Amin V, Wilson D, Rouse G, Upda S. Ultrasound image texture analysis for characterizing intramuscular fat content of live beef cattle. Ultrasonic Imaging 1998; 20: 191–205. 3. Neveu M, Dipanda A, Plantamp D, Diebold H. Segmentation d’images e´chographiques par analyse de texture. Innov Technol Biol Med 1996; 10: 413–428. 4. Corneloup G, Mosan J, Magnin IE. Bscan image segmentation by thresholding using co-occurrence matrix analysis. Pattern Recognition 1996; 29: 281–296.
90
5. Corneloup G, Mosan J, Magnin IE, Benoist P. Optimal co-occurrence matrix for automatic segmentation of ultrasonic images. Trait Signal 1996; 9: 309–323. 6. Pal SK, Pal NR. Segmentation using contrast and homogeneity measures. Pattern Recognition Lett 1987; 5: 293– 304. 7. Kass M, Witkin A, Terzopoulos D. Snakes: active contour model. Int J Comput Vision 1988; 10: 321–331. 8. Bossart PL, David D, Dinten JM, Chassery JM. Detection of regular boundaries in noisy and textured images: a multiscale active contour approach. Trait Signal 1997; 14: 209–225. 9. Cohen LD. On active contour models and balloons. Comput Vision Graphics, Image Understanding 1991; 53: 211–218. 10. Lam KM, Yan H. Fast greedy algorithm for active contour. Electron Lett 1994; 30: 21–23. 11. Lucassen GW, van der Sluys WLN, van Herk JJ, Nuijs AM. The effectiveness of massage treatment on cellulite as monitored by ultrasound imaging. Skin Res Technol 3: 154–160.
Address: Jean-Michel Lagarde Cerper, Institut de Recherche Pierre Fabre Hotel Dieu Saint-Jacques 2 rue de Viguerie 31025 Toulouse France Tel: 33 5 62 48 85 00 Fax: 33 5 62 48 85 99 e-mail:
[email protected]
