WO2002075660A2

WO2002075660A2 - Method and apparatus for real time contour tracking of objects on video images sequences

Info

Publication number: WO2002075660A2
Application number: PCT/IB2002/000844
Authority: WO
Inventors: Marcello Demi; Marco Paterni; Antonio Benassi; Vincenzo Gemignani
Original assignee: Consiglio Nazionale Delle Ricerche; Esaote S.P.A.
Priority date: 2001-03-21
Filing date: 2002-03-21
Publication date: 2002-09-26
Also published as: AU2002251394A8; AU2002251394A1; ITPI20010023A1; WO2002075660A3

Abstract

A method and an apparatus to track contours on sequences of video images in real-time. The method is mathematically efficient and is based on an operator which is robust against noise. Moreover, the new method does not introduce strong constraints on the shape of the object and can be used to track the contours of moving rigid objects as well as the contours of moving deformable objects. The contour of the structure under examination is determined automatically, in order to estimate the indices of interest of the organ associated to the contour and to display the results in real-time. To obtain real time performances an apparatus is provided based on a standard Personal Computer which was equipped with a DSP board. The PC provides a suitable environment for the development of a flexible and powerful user interface. The DSP board provides the image capture, the image display resources and the processing capability necessary to obtain the real time performances. Since both the PC and the DSP are programmable devices, the whole system is very flexible and is particularly suitable for the investigation of vascular mechanics in medical imaging.

Description

METHOD AND APPARATUS FOR REAL TIME CONTOUR TRACKING OF OBJECTS ON VIDEO IMAGES SEQUENCES DESCRIPTION Field of the invention

The present invention generally relates to a method for contour tracking on video images.

More precisely, the invention relates to a method for automatic contour tracking on sequences of grey level video images.

The images to treat may be of many types. In particular, but not exclusively, they are images of an organ, obtained with various techniques such as ultrasonic pulses, PET, SPECT, CAT, MR, etc., which can be anatomical images, or images of function, obtained by means of time sequences of anatomical views of a particular zone of the organ, or images of perfusion, obtained on the same organ after treatment of the patient with substances that highlight the perfusion in the organ. The bidimensional images may turn into volumetric images if acquired as slices of spatial sequences.

Description of the prior art

Image analysis, which is often based on grey level images, usually requires contour detection. Although expert operators can easily trace contours by means of a digitizer tablet or mouse, they cannot ensure the objectivity of the manually defined contours and therefore, the reliability of their quantitative analysis.

Moreover, due to the tedious and time consuming process of manual tracing, the studies which require the analysis of a great number of images cannot be accomplished. For example, some studies of the cardiac cycle from echocardiography require the analysis of a succession of frames taken from different points of view and consequently, they cannot be accomplished in clinical routine.

Adequate procedures to outline contours automatically are needed and many authors have addressed this issue as reflected by the large number of papers in literature. Usually, a contour detection algorithm requires: 1) filtering the images by means of an edge operator in order to provide an edge map which enhances the discontinuities between the grey levels of different structures, and 2) locating the points of the contour one is looking for through the edge map [1-3] . According to this approach, many authors have faced the problem and its main aspects which can be summarized as follows: - Each frame of a temporal or spatial sequence can be processed either by exploiting the frame data alone or by also exploiting the data given by the adjacent frames.

- Although subjectivity and operator time consumption are added, a human operator can be requested to give information about the desired structure.

- The optimum resolution to process locally every image point does not exist. In general, as edge operators have smaller masks greater accuracy of fine details is obtained, while with larger masks the global features are detected and the sensitivity to noise is reduced.

- The availability of a priori knowledge of the global shape of the objects present in the image sequence is another particular aspect of the problem. For example, even though the same anatomical structure shows different shapes in different patients, some morphological features are preserved.

- In spite of the large quantity of data to be dealt with, the entire procedure has to be performed in a reasonable amount of time. As regards the computational load, a strong reduction has been obtained by using a new edge operator. Standard isotropic edge operators, such as the Laplacian of Gaussian (LoG) and the magnitude of the Gradient of Gaussian (GoG) , are widely used in contour detection processes to enhance image luminance variations [3] . One of the advantages of using a LoG operator is economy of computation with respect to computing either directional derivatives or the GoG magnitude. Moreover, it is easier to locate the zero-crossing of the images which have been filtered by means of the LoG than to locate the peaks of the maps of the GoG magnitude. The LoG can be approximated by a difference of Gaussians (DoG) and of all the linear filters, the DoG is considered the most plausible early vision filter, since it best approximates the impulse response of the ganglion cells of the retina of both humans and other mammals [1] . The GoG magnitude, despite its extra computational cost, is adopted by a wide number of authors since it is more robust against noise than LoG [2] [4,5].

However, standard edge detectors optimally coping with monodi ensional discontinuities do not successfully deal with bidimensional discontinuities [5,6] nor do they deal with edge profiles which are different from the ideal one [7-9] . Therefore, many authors have addressed the problem of detecting and analyzing these particular discontinuities [9-14] .

In particular, in [15] a filter that can provide ridges at edges and ridge local maxima at junctions and corners is held to be the simplest way to overcome the problem. Such a filter can be obtained from the generalization of the first absolute central moment. This filter belongs to the wide class of moments which includes mean, variance, "skewness" and "kurtosis" [16] .

However, unlike these moments, the absolute moments have not been investigated in the past because of the mathematical difficulties introduced by the absolute value brackets [17] . The first order absolute moment has become a popular operator only in recent years [17] as the subject of robust estimation [18] has become increasingly more important .

Statistical measures, such as mean, variance, skewness and kurtosis are also used in literature to describe the spatial features of a region of the image [19,20]. Mean is also used to reduce noise even though the use of the mean filter is limited to the cases where the noise distribution well approximates a normal distribution since the averaging operation is not robust in the presence of noise with heavy-tailed distribution [21] . Moreover, the mean filter tends to blur edges present in the images. A filter which reduces noise and does not possess these two drawbacks is the median filter. The median and the mean deviation (first absolute central moment) are largely used in robust statistics when those cases for which the normal model would be a bad approximation are considered. However, unlike the median filter which is largely used in image processing applications [21] , the mean deviation is not used in this context.

More generally, few authors have referred to the edge detection properties of the central and absolute central moments [20] [22,23]. All the central moments with an odd order enhance the grey level discontinuities with a zero- crossing while all the absolute central moments enhance the grey level discontinuities with a ridge [24] . However, of all the central and absolute central moments, only the first moments should be used with this aim since they are the moments that involve the lowest power of the input data. Higher moments involving higher power of the input data are almost always less robust to corrupting noise [17] .

Another method is known, i.e. the Block Matching Algorithm (BMA) to estimate the movement of an object. The BMA assumes that the grey level distribution of a block of pixels varies slowly during an image sequence. According to this assumption an object can be automatically located in a sequence of images by looking for similar blocks of pixels on every pair of consecutive images. In US5.999.651, US6.137.913, and US5.946.041 the BMA was used to estimate the movement of an image feature on two consecutive frames. In particular, in US5.999.651, first a global motion of the object is computed by using a prediction algorithm and then the local motion is estimated by a modified BMA. Both the patents US6.137.913 and US5.946.041 use the BMA to obtain motion information of the contour under examination by starting from information derived from the previous image of the sequence.

Other patents adopt methods based on threshold algorithms. One example can be found in US5.289.373 where the problem of tracking the catheter guide in real-time on fluoroscopic images is considered.

US 5.943.441 and US 5.862.245 provide a method to automatically locate a contour in a single image by starting from a given point. In US 5.943.441 the contour is progressively located by finding a sequence of one- dimensional edge positions where every edge position is determined by processing a set of pixels arranged along an imaginary line. In US 5.862.245 an iterative process deforms an active contour until a first token of the true contour is obtained. The token is then extended by straight line segments. Only US 5.943.441 and US 5.289.373 introduce apparatuses based on custom circuitry with the aim of executing their algorithms in real time. The other cited patent documents do not deal with the problem of obtaining real-time performances. Summary of the invention

It is an object of the present invention to provide a contour tracking procedure on video images arranged as a sequence of frames, such as a plurality of parallel slices of an organ, that reduces operator intervention to tracing a rough contour on the first frame of the sequence and exploits the reciprocal influence of the data of adjacent frames by using the contours outlined in a previous frame as a starting contour for the next frame.

It is another object of the present invention to provide a method to track contours on sequences of video images in real-time that:

- is mathematically efficient;

- is based on a operator which is robust against noise;

- does not introduce strong constraints on the shape of the object;

- can be used to track the contours of moving rigid objects as well as the contours of moving deformable objects . It is still another object of the present invention to provide a contour tracking procedure of video images in which:

- noise sensitivity is reduced without loss of accuracy, - a multiresolution approach is possible to localise the boundary points;

- a strong reduction of the computational load is obtained with respect to the prior art.

According to the invention, these objects are achieved by an automatic contour tracking method on sequences of grey level video images using a new edge operator defined as the mass centre of the grey level variability.

The method according to the invention provides the steps of: - - On a first frame of the sequence, manually or semi-automatically tracing a starting contour following an edge discontinuity on a grey level map of an image, the starting contour being defined by selecting a plurality of pixels; - computing the mass centre of the grey level variability at each pixel of the starting contour;

- repeating the previous step for each mass centre for several iterations up to convergence to a final mass centre; - replacing each pixel of the starting contour with each final mass centre thus obtaining the final contour on the first frame;

- using the final contour of a frame as a starting contour on the next frame of the sequence and repeating the above steps to obtain the final contour on the next frame .

The step of computing the mass centre of the grey level variability at each pixel of the starting contour comprises the further steps of: - Defining two domains ®χ and Θ₂ surrounding the point of the starting contour;

- associating to the first domain Θi a first weight function i and calculating the mean value μ of the grey level map on ©i;

- associating to the second domain Θ₂ a second weight function w₂ and calculating on Θ₂ the first absolute central moment where the considered mean value is the mean value μ computed on Θi;

- calculating the vector b by adding the absolute differences between the mean value μ and the grey levels of the pixels pi belonging to Θ₂ multiplied by the vector joining pi to the centre of Θ₂ and divided by the first absolute central moment;

- associating the point indicated by the vector b to the mass centre of the grey level variability which is a point closer to the discontinuity than the starting point, independently of the distance between the latter and the discontinuity.

Advantageously, the weight functions are chosen among Gaussian functions.

- Preferably, two Gaussian functions are used as weight functions that are normalized on circular domains having radii that are two times the apertures of the Gaussians, a configuration σι= σ₂ being adopted to iteratively locate the mass centre of the grey level variability.

Advantageously, the step of replacing the points of the starting contour with the mass centres allows the application of additional constraints.

The application of additional constraints may comprise the step of defining the approximate contour as a curve interpolating N predetermined knots.

The curve is preferably chosen among: a polygonal line and a cubic spline.

According to another aspect of the invention, an apparatus for contour tracking on video images arranged as a sequence of frames comprises a DSP board interacting with a Personal Computer, the DSP capturing, processing and displaying the video signals in real-time and interfacing with the PC, the apparatus being characterised in the edge operator that the PC uses defined as the mass centre of the grey level variability .

The video images are anatomical, or functional or perfusion images of an element, obtained with suitable medical imaging techniques chosen for example among: RX, angiography, ultrasonic pulses, PET, SPECT, CAT, MR.

In the apparatus two monitors are preferably provided for displaying the video sequence and the user interface, respectively; a mouse and a keyboard are available as user interface tools; two input buffers are used for storing the input images in an alternate way.

Brief description of the drawings

Further characteristics and advantages of the method for treating anatomical images according to the present invention will be made clearer with the following description of an embodiment thereof, exemplifying but not limitative, with reference to attached drawings wherein:

- figure 1 shows a circular domain at a point p which contains a grey level discontinuity at a distance d;

- figure 2 shows an approximate starting contour Cs and the final contour Ci of an object of interest on the first frame of an image sequence;

- figure 3 shows the determination steps, starting from bidimensional images, of the contour of an organ;

- figure 4 shows a diagrammatical view of an apparatus for carrying out the invention comprising a standard

Personal Computer equipped with a DSP board, as well as two monitors and the user interface, respectively. A mouse and a keyboard are available as tools for the user interface; - figure 5 shows in a schematic way the data path in the apparatus of figure 4;

- figure 6 shows a block diagram of the apparatus of figure 4 interfaced to an ultrasound system equipped with a transesophageal probe for recording the image sequences of the aorta;

- figure 7 shows a diagram of the software architecture resident in the apparatus of figure 6; - figure 8 shows the mass centre vectors bi,j of 2N+1 consecutive points taken around knots Ki in order to obtain the knots of a new contour which better approximates the contour of the aorta. Description of a preferred embodiment

In grey level medical images the objects to detect are the discontinuities, that are often associated to boundary walls of organs.

To introduce the mathematical operator mass centre of the grey level variability, according to the present invention, a grey level map f (n,m) of an image is considered. Normally, in a grey level map each pixel may be associated to 256 grey levels, i.e. from 0 to 255.

The first absolute central moment used for determining the mass centre of the grey level variability is a statistical filter which measures the variability of the grey levels of the image with respect to a local mean. Given a point p, the spatial distribution of the variability of the grey levels with respect to the local mean computed at point p can be seen as a mass density function which associates a mass value to every pixel surrounding p. In this case the first absolute central moment is the total mass of the variability of the grey levels at point p. The mass centre of the grey level variability with respect to the local mean can be defined as a vector b. Vector b when computed at a point p close to a grey level discontinuity locates a point which is closer to the discontinuity than p, independently of the distance between the latter and the discontinuity. Therefore, the properties of the mass centre of the grey level variability is used according to the invention to develop a new contour tracking procedure.

Let ©x and Θ₂ be two circular domains defined as:

®i = {(k,l) e z² ^k² + l² ≤ r_i} wherein Z represents the integer numbers, the mean value of f (n,m) on the circular domain ©i at a point P=(_I,_Ώ) is /i (P) = ∑(_k, l) e Θi ft" ^{~ k}'^{m ~ 7})^W^' ^{l> r}ι wherein w (k, l , ∑ι) is a weight function with a unitary summation on the domain ®_λ . Then, every pixel of the circular domain Θ₂ is associated to a mass h (~p , k, l) so that:

wherein w (k, l , r_∑) is a weight function with a unitary summation on the domain Θ₂. The function h (p , , 1) represents the spatial distribution of the variability of the grey levels of the domain ©₂ with respect to the local mean fι (p) computed at point p. The mass centre of the function h (ιp , k_r l) is computed as:

b(p) = ∑∑₍*,_{Q eθ}^P^) ^{∑∑(k )}'^{@2 /)r lf} ∑∑⁽^<^^(pΛ)≠°

wherein T is a discrete vector with components (-k,-l) .

Vector b joins point p to the mass centre p' of the variability of the grey levels of the domain ©₂.

With reference to figure 1, if a grey level discontinuity is present and a point p is distant d from the discontinuity as shown in Fig. 1, if d<r₂ and if the right configuration ( wχ , w₂) of the operator is chosen, then the mass centre p' of the grey level variability is closer to the discontinuity than p, independently of the distance d. Therefore, the nearest point pi of the discontinuity can be located by iteratively computing the mass centre of the grey level variability. When any new iteration occurs, the starting point is the mass centre, determined by means of the previous iteration. The procedure converges quickly and few iterations are necessary to reach the discontinuity. The mass centre of the grey level variability can be then used to locate a discontinuity by starting from an approximated contour.

This approach is different from the classical methods based on the derivative operators. The classical methods locate the discontinuity by searching for local maxima or zero-crossings in a neighborhood of the starting contour which, however, can be rather large. According to the invention, the mass centre of grey level variability "jumps" to the discontinuity in just a few steps and, consequently, the computational cost of our method is generally less than when methods based on derivative filters are used.

With reference to Fig.2, Cs is an approximate starting contour and CI is the contour of the object of interest on the first frame of an image sequence. If the mass centre of the grey level variability is iteratively computed at every point of Cs, then all the points of Cs move to their respective nearest points of CI. Once CI is computed, it is used as the approximate starting contour to compute C2 in the second frame and so on. In conclusion, given an approximate starting contour on the first frame of a sequence, the edge operator according to the invention can be used to track automatically the contour of the object of interest through the sequence. However, the procedure may require some additional constraints. In fact, when the mass centre of the grey level variability is computed at a point of the image, the nearest point of the nearest discontinuity is detected. If a second discontinuity is close to the object, then some of the points of the approximate contour can move to the second discontinuity instead of moving to the contour of the object under examination. In order to overcome this localization failure, additional constraints may be added to the shape of the contour. For example, the contour of the object is described as a curve interpolating N knots Ki (four knots in fig. 8) since the number of knots and the type of interpolation set restrictions on the shape of the contour. In case of need, however, other constraints can be added to the location and to the movement of the knots.

The method according to the invention, comprising the addition of constraints, is summarized in Fig. 3: - on a first frame of the sequence (1), an approximate starting contour is manually or semi-automatically traced (2);

- then the mass centre of the grey level variability is computed (3) at the points of the approximate contour; - the additional constraints (4) are applied and the contour of the object is obtained;

- the obtained contour is used as an approximate starting contour on the next frame (5) of the sequence.

The computational cost of the procedure is primarily due to the iterative computation of the mass centre of the grey level variability. However, the algorithm and the number of steps are the same for all the points and every mass centre is computed separately at every single point of the contour. Therefore, an efficient parallel implementation of the contour tracking procedure is possible.

The contour tracking procedure according to the invention runs in real-time in an apparatus which is based on the interaction between a DSP board and a Personal Computer. The DSP board provides the resources necessary to capture, process and display the video signal in realtime while the PC acts as a flexible and powerful user interface. This sort of architecture is completely programmable and provides a very versatile apparatus which can be adapted to several contour tracking applications.

With reference to figure 4, block 11 is a standard Personal Computer equipped with a DSP board. Two monitors 12 and 13 are used to display the video sequence and the user interface, respectively. A mouse 14 and a keyboard 15 are available as tools for the user interface.

Fig. 5 shows the data path in a schematic way. The apparatus accepts both NTSC and PAL video standards 21 as input signals. The images are captured (22) and temporarily stored in the memory of the DSP board. Two input frame buffers 23 are present and the input images are alternately stored in the two buffers . Once an image is entirely captured in memory, the DSP 24 locates the contour of the object of interest automatically by using the contour determined on the previous frame as an approximate starting contour. Afterwards, the contour is graphically superimposed on the image and the frame data are transferred to the image display hardware 25. Therefore, when the processing activity is started, the monitor 26 displays the video sequence with a superimposed outlined contour. The contour data are also transferred in real-time from the DSP to the PC 27 and the contour is graphically displayed in real time on the monitor 28 which is used as a user interface. Example

The method according to the invention has been used to analyse the cross sectional area changes of the aorta during pharmacological intervention. The aorta, recorded by transoesophageal echocardiography (TEE) , was observed before and after i.v. injection of 2 mg of Isosorbide Dinitrate. The vasodilatation due to the injection of the Isosorbide Dinitrate provides important information about vascular mechanics. The proposed system acquires the video signal, allows the operator to trace an approximate starting contour of the aorta on the first frame of the sequence, locates the contours of the aorta on every subsequent frame of the sequence, computes the cross sectional area of the aorta, and displays both the data and the images in real time.

In the apparatus an ultrasound system equipped with a transesophageal probe is used to record the image sequences of the aorta. The contour tracking procedure based on the mass centre of the grey level variability was developed on a TMS320C80 Software Development Board (SDB) of Texas Instruments.

The SDB is a PCI board designed as a tool to develop real time audio/video applications. It provides an audio interface, video capture and display resources of up to 30 frames/sec, on board memory, and software tools to develop codes on the C80. The SDB can communicate with the Personal Computer by means of a client-server protocol. The heart of the SDB is the Texas Instruments TMS320C80 single-chip multiprocessor. It is a high performance and highly integrated digital signal processor especially designed for use in image processing and in both two- dimensional and three-dimensional graphics. It is capable of performing the equivalent of over two billion RISC-like operations per second. The SDB communicates with the PC either to receive commands or to transfer the coordinates of the points of the outlined contour.

The output video signal of the ultrasound system is grabbed by the SDB and converted into digital images with a resolution of 512X512 pixels and 8 bit per pixel. Afterwards, the TMS320C80 processes the image by iteratively computing the mass centre of the grey level variability at the points of the starting contour. Finally, the computed contour is superimposed on the image and at the same time transferred to the PC by the PCI bus. The image with the superimposed contour is instead converted to a VGA signal and displayed on the monitor 31. A switch sends the video output signal of the ultrasound system or the video output signal of the SDB to the VGA monitor when either standard clinical applications or vascular studies during pharmacological intervention, respectively, are performed. Moreover, once the computed contour has been transferred to the PC, opportune procedures compute and graphically display the cross sectional area changes of the aorta.

The preferred procedure is the following. The system can be split into two separated operative units since two distinct procedures run simultaneously: the contour localization procedure (CLP) which runs on the SDB and the user interface procedure (UIP) which runs on the PC. At the system startup the PC loads the software on the SDB and sets all the parameters to the default values. When the system is running the UIP verifies which of the following events happens:

Command. This happens when the user interacts with the system in order to trace a new starting contour, to set new parameters, to delete a wrong contour or to start or stop the localization procedure. In this case the UIP communicates the user's data, parameters and commands to the SDB. The following functions are available: manually trace the approximate starting contour, delete the approximate starting contour, locate/do not locate the contour, display the plot of the cross sectional area changes, display the plot of the diameter changes, and store the data set (contours, areas and diameters) .

Contour. This happens when the SDB sends the computed contour to the PC. In this case the UIP executes the following steps: displays the contour, computes the cross sectional area and the diameter, and plots as a graph the values of area and diameter.

After the start-up the SDB runs the CLP that verifies which of the following events happens:

New frame. This happens when the frame grabber captures a new digital image. In this case the CLP locates the new aortic contour by starting from the aortic contour determined on the previous image and sends the new contour to the PC.

Command. This happens when the UIP sends commands and/or data to the SDB. In this case the CLP performs the command received from the user.

The procedure was quantitatively evaluated by using the linear regression of the cross-sectional area of the aorta which had been detected by the computer and provided by an expert. We used 20 pharmacological studies with images of diverse quality. Thirty seconds of every study were analyzed and the cross sectional area of the aorta was automatically evaluated both in basal conditions and after vasodilator injection. Afterwards, the contours of the aorta were traced manually on 25 consecutive frames of every study by using the tools of the ultrasound system. We then compared the cross sectional areas of the aorta derived from these contours to the cross sectional areas provided on the same frames by the automatic procedure.

Let i be an index which varies between 1 and Q and let j be an index which varies between 1 and 2N+1 , the starting contour was approximated with a spline interpolating Q knots and 2N equidistant points were chosen between each pair of adjacent knots as depicted in Fig. 8. Subsequently the vectors b±._j of the 2N+1 consecutive points taken around each knot Ki were computed. Then, the mean values b± of the 2N+1 vectors bχ,_j were computed and associated to the respective knots Ki. Once the vectors b were computed they provided the knots of a new contour which better approximated the contour of the aorta we were looking for. The procedure was repeated iteratively and the final contour was located in three iterations. A cubic spline, interpolating 4 equispaced knots and a value of N equal to 2 were used.

Two Gaussian functions g (k _r l , σ ) were used for the weight functions w (k, l , r±) . The Gaussian functions were normalized on two circular domains which had a radius that was equal to two times the apertures of the Gaussian

{ χ=2σi) and the configuration σι= σ₂ was adopted to iteratively locate the mass centre of the grey level variability.

The linear regression analysis shows a good correlation between the cross sectional areas of the aorta which derived from the manually traced contours and the cross sectional areas provided on the same frames by the automatic procedure: the slope varies between 0.95 and 0.98, the intercept varies between 9.275 and 12.221, the regression coefficient varies between 0.94 and 0.98. If either the number of the knots or the value of N increases the performances of the procedure do not increase significantly. The system fails only in the case of abrupt movements of the ultrasound window and in the case of inadequate echocardiographic imaging.

The experimental results show that the procedure according to the invention is more robust against noise than the procedures based on standard edge detectors such as the GoG and the LoG. Moreover, the computational cost of procedures based on the mass centre of the grey level variability is less than when methods based on GoG or DoG filters are used.

The foregoing description of a specific embodiment will so fully reveal the invention according to the conceptual point of view, so that others, by applying current knowledge, will be able to modify and/or adapt for various applications such an embodiment without further research and without parting from the invention, and it is therefore to be understood that such adaptations and modifications will have to be considered as equivalent to the specific embodiment. The means and the materials to realise the different functions described herein could have a different nature without, for this reason, departing from the field of the invention. It is to be understood that the phraseology or terminology employed herein is for the purpose of description and not of limitation.

REFERENCES 1. D. Marr, Vision, San Francisco: W.H. Freeman, 1982.

2. J. Canny, A computational approach to edge detection, IEEE Trans . Pattern Anal . Machine Intell . , vol. PAMI- 8, 1986, pp. 679-698.

3. V. Torre and T. Poggio, On edge detection, IEEE Trans . Pattern Anal . Machine Intell . , vol. PAMI-8,

1986, pp. 147-163.

4. D.H. Ballard, CM. Brown, Computer Vision, Prentice- Hall, 1982.

5. E. De Micheli, B. Caprile, P. Ottonello and V. Torre, Localization and Noise In Edge Detection, IEEE Trans .

Pattern Anal . Machine Intell . , vol. 11, 1989, pp.1106-1117.

6. V.Berzins, Accuracy of Laplacian edge detectors, Comput . Grap . Image Processing, vol.27, pp. 195-201, 1984.

7. M.C.Morrone and R.A.Owens, Feature detection from local energy, Pattern Recognition Letters, vol.6,

1987, pp.303-313. 8. R.Owens, S.Venkatesh, J.Ross, Edge detection is a projection, Pattern Recognition Letters, vol.9, 1989, pp.233-244.

9. L.Rosenthaler, F.Heitger, O.Kubler and R.von of the Heydt, Detection of General Edge and Keypoints,

Lecture notes in computer science: Computer Vision, Springer-Verlag, 1992, pp.78-86.

10. L. Kitchen and A.Rosenfeld, Grey-level corner detection, Pattern Recognition Letters, vol.l, 1982, pp.95-1002.

11. K.Rangarajan, M.Shah, and D.Van Brackle, Optimal Corner Detector, Comput . Vision Graph. Image Processing, vol.48, 1989, pp.230-245.

12. K.Rohr, Modelling and identification of characteristic intensity variation, Image and Vision

Computing, vol.10, 1992, pp.66-76.

13. J.A. Noble, Finding half boundary and junctions in images, Image and Vision Computing, vol.10, 1992, pp.219-232. 14. P.Perona, Steerable-Scalable Kernels for Edge Detection and Junction Analysis, Lecture notes in computer science: Computer Vision, Springer-Verlag, 1992, pp.3-18.

15. M. Demi, in the approach to automatic contour detection from image sequences: an application to ventriculographic images, Comput . Biomed. Res . 27, 1994, pp. 157-177.

16. A. Papoulis, Probability, Random Variables , and Stochastic Process, Mc Graw-Hill, in the York, 1965. 17. W. H. Press, B. P. Flannery, S. A. Teukolsky, W. T. Vetterling, Numerical Recipes, Cambridge University Press, 1991. 18. P.J. Huber, Robust Statistics, in the York, Wiley, 1981. 19. W. K. Pratt, Digital image processing, in the York, Wiley, 1991. 20. A. K. Jain, Fundamentals of digi tal image processing, Prentce-Hall, 1989. R. Chellappa, Digital image processing, Los Alamitos CA, Computer Society press, 1992 B. Jahne, Digi tal Image Processing, Springer-Verlag, 1997. T. Yoo, Image Geometry Through Multiscale Statistics, Ph . D. Disserta tion, University of North Carolina at Chapel Hill, Department of Computer Science, 1996.

Claims

1. Method for tracking contours on sequences of images arranged as a sequence of grey level frames, characterised in that an edge operator defined as the mass centre of the grey level variability is used.

2. Method according to claim 1 characterised in that it comprises the steps of:

- on a first frame of the sequence, manually or semi- automatically tracing a starting contour (Cs) following an edge discontinuity on a grey level map f(n,m) of an image, said starting contour (Cs) being defined by selecting a plurality of pixels;

- computing the mass centre (p' ) of the grey level variability at each pixel (p) of said starting contour (Cs);

- repeating the previous steps for each mass centre (p' ) for several iterations up to convergence to a final mass centre (pi) ;

- replacing the points (p) of said starting contour with said final mass centres (pi) thus obtaining the final contour (Ci) on the first frame;

- using said final contour and any subsequent final contour as a starting contour (Ci_ι) on a next frame (i) of the sequence and repeating the above steps to obtain the final contour (Ci) on the next frame.

3. Method according to claim 2 characterised in that said step of computing the mass centre (p' , pi) of the grey level variability at each pixel (p) of said starting contour (Cs) comprises the further steps of: - defining two domains ©i and Θ₂ surrounding the point p of the starting contour;

- associating to the domain ©₂ a weight function w₂ and calculating on Θ₂ the first absolute central moment where the considered mean value is the mean value μ computed on ©i; — calculating a vector b by adding the absolute differences between the mean value μ and the grey levels of the pixels pi belonging to Θ₂ multiplied by the vector joining p_± to the centre p of ©₂ and divided by the first absolute central moment;

4. Method according to claim 3 characterised in that, said weight functions (w) are preferably chosen among Gaussian functions (g) .

5. Method according to claim 4 characterised in that, two Gaussian functions are used as weight functions that are normalized on circular domains having a radius that is two times the apertures of the Gaussians { rχ=2σ_±) , a configuration σι= σ₂ being adopted to iteratively locate said mass centre of the grey level variability.

6. Method according to claim 3 characterised in that, said step of replacing said points (p) of said starting contour (Cs) with said mass centres (p' , pi) allows the application of additional constraints (K) .

7. Method according to claim 6 characterised in that, said application of additional constraints (K) comprises the step of defining the approximate contour as a curve interpolating N predetermined knots (K) .

8. Method according to claim 7 characterised in that, said curve is preferably chosen among: a polygonal line and a cubic spline.

9. Method according to claim 7 characterised in that, said video images are anatomical, or functional or perfusion images of an element, obtained with suitable medical imaging techniques chosen among: RX, angiography, ultrasonic pulses, PET, SPECT, CAT, MR.

10.An apparatus for contour tracking on video images arranged as a sequence of frames comprises a DSP board interacting with a Personal Computer, said DSP capturing, processing and displaying the video signals in real-time and interfacing with said PC, characterised in that said apparatus uses an edge operator defined as the mass centre of the grey level variability .

11.Apparatus according to claim 10, wherein said edge operator is implemented in said apparatus according to claims 2 to 9.

12.Apparatus according to claim 10, characterised in that the following are provided:

- two monitors for displaying the video sequence and the user interface, respectively; - a mouse and a keyboard available as user interface tools; - two input buffers are used for storing the input images in an alternate way.

13. Apparatus according to claim 10, characterised in that said DSP board automatically locates the contour of the object of interest using the contour of the previous frame as approximate starting contour.