WO2003030101A2

WO2003030101A2 - Detection of vessels in an image

Info

Publication number: WO2003030101A2
Application number: PCT/DK2002/000662
Authority: WO
Inventors: Ebbe Sørensen; Johan Doré HANSEN; Michael Grunkin; Niels Vaever Hartvig; Jannik Godt; Per Rønsholt ANDRESEN; Soffia Björk SMITH
Original assignee: Retinalyze Danmark A/S
Priority date: 2001-10-03
Filing date: 2002-10-03
Publication date: 2003-04-10
Also published as: WO2003030101A3

Abstract

The present invention relates to a method for detecting vessels in an image and a system therefor, wherein said image may be any image comprising vessels, in particular an image from medical image diagnostics, and more particularly an ocular fundus image. The method includes tracking vessel segments from starting points, wherein the tracking process correlates identified vessel segments with estimated values for the vessel segments. Also the invention provides a validation of tracked vessels when the tracking process stops due to stop criteria.

Description

Detection of vessels in an image

The present invention relates to a method for detecting vessels in an image and a system therefor, wherein said image may be any image comprising vessels, in parti- cular an image from medical image diagnostics, and more particularly an ocular fundus image.

Background

Broadly, two approaches exist for vasculature analysis. One approach, hereafter referred to as the pixel-processing approach, works by filtering or segmentation, followed by thinning and branch point analysis. These methods usually require the processing of every image pixel, and thus they scale poorly with image size. Besides, they often fail to distinguish between vessels and other features, such as other anatomical structures or pathologies.

The second approach is referred to as vessel tracking. These methods work by first locating an initial point and then exploiting local image properties to trace the vasculature recursively. They only process pixels close to the vasculature, avoiding the processing of every image pixel, and so are appropriately referred to as exploratory algorithms. They have several properties that make them attractive relative to pixel- processing approaches; they scale well with image size, and can be made highly adaptive, while being robust, efficient, and accurate.

Fundus image analysis presents several challenges, such as high image variability, the need for reliable processing in the face of nonideal imaging conditions and short computation deadlines. Large variability is observed between different patients - even if healthy, with the situation worsening when pathologies exist. For the same patient, variability is observed under differing imaging conditions and during the course of a treatment or simply a long period of time. Besides, fundus images are often characterized by having a limited quality, being subject to improper illumination, glare, fadeout, loss of focus and artifacts arising from reflection, refraction, and dispersion. Automatic extraction and analyzation of the vascular tree of fundus images is an important task in fundus image analysis for several reasons. First of all the vascular tree is the most prominent feature of the retina, and it is present regardless of health condition. This makes the vascular tree an obvious basis for automated registration and montage synthesis algorithms. Besides, the task of automatic and robust localization of the optic nerve head and fovea, as well as the task of automatic classification of veins and arteries in fundus images may very well rely on a proper extraction of the vascular tree. Another example is the task of automatically detecting lesions which in many cases resemble the blood vessels. A properly extracted vessel tree may be a valuable tool in disqualifying false positive responses produced by such an algorithm, thus increasing its specificity. Finally the vessels often display various pathological manifestasions themselves, such as increased tortuosity, abnormal caliber changes and deproliferation. An automatic vessel tracking algorithm would be the obvious basis for analysis of these phenomena as well.

Diabetes is the leading cause of blindness in working age adults. It is a disease that, among its many symptoms, includes a progressive impairment of the peripheral vascular system. These changes in the vasculature of the retina cause progressive vision impairment and eventually complete loss of sight. The tragedy of diabetic reti- nopathy is that in the vast majority of cases, blindness is preventable by early diagnosis and treatment, but screening programs that could provide early detection are not widespread.

Promising techniques for early detection of diabetic retinopathy presently exist. Re- searchers have found that retinopathy is preceded by visibly detectable changes in blood flow through the retina. Diagnostic techniques now exist that grade and classify diabetic retinopathy, and together with a series of retinal images taken at different times, these provide a methodology for the early detection of degeneration. Various medical, surgical and dietary interventions may then prevent the disease from progressing to blindness.

Despite the existing techniques for preventing diabetic blindness, only a small fraction of the afflicted population receives timely and proper care, and significant barriers separate most patients from state-of-the art diabetes eye care. There are a lim- ited number of ophthalmologists trained to evaluate retinopathy, and most are lo- cated in population centers. Many patients cannot afford the costs or the time for travel to a specialist. Additionally, cultural and language barriers often prevent elderly, rural and ethnic minority patients from seeking proper care. Moreover, because diabetes is a persistent disease and diabetic retinopathy is a degenerative disease, an afflicted patient requires lifelong disease management, including periodic examinations to monitor and record the condition of the retina, and sustained attention on the part of the patient to medical or behavioral guidelines. Such a sustained level of personal responsibility requires a high degree of motivation, and lifelong disease management can be a significant lifestyle burden. These factors increase the likeli- hood that the patient will, at least at some point, fail to receive proper disease management, often with catastrophic consequences.

Accordingly, it would be desirable to implement more widespread screening for retinal degeneration or pathology, and to positively address the financial, social and cultural barriers to implementation of such screening. It would also be desirable to improve the efficiency and quality of retinal evaluation.

Hence, a precise knowledge of both localisation and orientations of the strucutures of the fundus is important, including the localisation of the vessels. Currently, exami- nation of fundus images is carried out principally by a clinician examining each image "manually". This is not only very time-consuming, since even an experienced clinician can take several minutes to assess a single image, but is also prone to error since there can be inconsistencies between the way in which different clinicians assess a given image. It is therefore desirable to provide ways of automating the process of the analysis of fundus images, using computerised image analysis, so as to provide at least preliminary screening information and also as an aid to diagnosis to assist the clinician in the analysis of difficult cases.

Next, it is generally desirable to provide a method of accurately determining, using computerised image analysis techniques, the position of both the papilla (the point of exit of the optic nerve) and the fovea (the region at the centre of the retina, where the retina is most sensitive to light), as well as vessels of the fundus. Summary of the invention

The present invention relates to a method for detecting vessels in an image, wherein said image comprises a plurality of vessels. The image may be an image of any subject comprising vessels. In particular the method relates to image diagnostics in medicin, such as X-rays, scanning images, photos, magnetic nuclear radiation scanning, CT scannings, as well as other images comprising vessels. The images are normally 2-dimensional or 3-dimensional and mostly comprise a network of vessels, normally presented as vessel trees; however, the vessels may also be archa- des of vessels. The images may be from any part of the body, in particular the present invention relates to images of the ocular fundus wherein the vessels may be seen directly by viewing through the pupil of the eye. In order to be able to make automatic detection of various structures in fundus images a reliable method of detecting the vessels in fundus images is necessary. Furthermore, the method should be robust in the sense that it should be applicable to a wide variety of images independent of illumination, presence of symptoms of diseases and/or artefacts of the image.

Accordingly, the present invention relates to a method for tracking at least one ves- sel in an image, said image comprising a plurality of vessels, comprising

a) establishing at least one starting point representative for a vessel in the image,

b) selecting at least one starting point representative for a vessel, detecting at least two characteristics of a vessel segment comprising the starting point, optionally determining a corrected starting point related to the vessel segment, and determining at least one vessel state, said vessel state comprising information of at least one of the characteristics detected,

c) from an, optionally corrected, starting point identifying a neighbouring vessel segment by the steps of:

d) detecting corresponding characteristics for the neighbouring vessel segment, e) providing estimates for the characteristics for the neighbouring vessel segment, and

f) for at least one characteristic individually weighting the detected characteristic with the estimated characteristic thereby obtaining a validated characteristic,

g) determining a vessel state for the neighbouring vessel segment, said vessel state comprising information of at least one of the detected characteristics and/or the validated characteristics,

h) optionally determining a corrected starting point related to the neighbouring vessel segment,

i) repeating steps c) to h) until a stop criterion has been fulfilled, thereby tracking a vessel comprising the vessel segments identified,

j) optionally repeating steps b) to i) until all starting points from step a) have been selected.

By the method at least one of the characteristics of each vessel segment is weighted with a predetermined value, preferably, at least two are weighted individually, whereby the estimated characteristics are filtered with predetermined characteristics, for example by a Kalman filtering method. By this method an estimated vessel segment gives a prescribed weight to the neighbouring vessel segment.

Another aspect of the invention relates to a method for tracking vessels in an image, wherein the tracking procedure comprises

a) identifying a consecutive sequence of vessel segments until a stop criterion is fulfilled, and for each vessel segment determining a vessel state having a vessel state significance, optionally additionally determining a filtered vessel state,

b) selecting a consecutive sequence of vessel segments, wherein each vessel segment has a vessel state significance and/or a filtered vessel state signifi- cance above a predetermined threshold, c) identifying the consecutive sequence of vessel segments selected in step b) as a vessel.

The tracking algorithms presented herein utilizes a number of characteristics associated to the retinal vessels: They can be assumed to have lower reflectance than their local background, they have well-defined edges, and the intensity within them varies smoothly. Furthermore, the vessels are locally continuous; changes in position, direction, caliber and intensity between branching points are smooth. Another important heuristic is the fact that the vascular tree is in general connected. However, an image may present a partial view, so the vessels are not all expected to be connected in a partial image.

When recording fundus images, at least 4 images are normally recorded from each fundus, representing different regions of the fundus. In order to examine the fundus region properly, registering or mounting of the images in a continuous manner with respect to the structures in the image, such as for example by arranging the images so that the vessels correctly continue in the images. Accordingly, the present invention relates to a method for registering at least two different fundus images of the same fundus, comprising detecting the vessels in said images by a method as defined above, and orienting the images with respect to the vessels.

Also, the detection of vessels may be used for detecting the optic nerve head in the image, optionally as an iterative method. Thus, the present invention relates to a method for detecting vessels in a fundus image, comprising

a) estimating the localisation of the vessels by a method as defined above,

b) estimating the optic nerve head based on the localisation of the vessels,

c) optionally repeating steps a) and b) at least once.

An important aspect of the present invention is the application of the method in a method for assessing the presence or absence of lesions in a fundus image. In such an application the detection of the vessels is used to for example mask the vessels in order to avoid false positive lesions very likely detected in the neighbourhood of the optic nerve head, or to merely adjust lesions detected with the vessels. Thus, the invention also relates to a method for assessing the presence or absence of lesions in a fundus image, comprising

a) detecting the localisation of the vessels in the image if present,

b) masking a region comprising the vessels, and estimating the presence or absence of lesions in the remaining image, or

c) estimating the presence or absence of lesions in the remaining image, and adjusting the lesions with respect to the vessels detected.

Any pathological conditions having any relation to the retina may be detected by use of a method wherein the vessels is detected as described herein. In particular pathological conditions having symptoms relating to the vessels themself, such as micro- aneurysms, turtuosity, dilations, constrictions and hemorrhages may be detected.

Furthermore, the invention relates to a system for carrying out the methods accord- ing to the invention, such as a system for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

an image acquisition unit,

a computer processor, and

a memory coupled to the computer processor, said memory storing said image, and storing

a) an algorithm for establishing at least one starting point representative for a vessel in the image,

b) an algorithm for selecting at least one starting point representative for a vessel, detecting at least two characteristics of a vessel segment comprising the starting point, optionally determining a corrected starting point related to the vessel seg- ment, and determining at least one vessel state, said vessel state comprising information of at least one of the characteristics detected,

c) an algorithm for, from an, optionally corrected, starting point, identifying a neigh- bouring vessel segment by the steps of:

d) detecting corresponding characteristics for the neighbouring vessel segment,

e) providing estimates for the characteristics for the neighbouring vessel segment, and

h) optionally determining a corrected starting point related to the neighbouring ves- sel segment,

i) an algorithm for repeating steps c) to h) until a stop criterion has been fulfilled, thereby tracking a vessel comprising the vessel segments identified,

j) an algorithm for optionally repeating steps b) to i) until all starting points from step a) have been selected,

as well as

a system for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

an image acquisition unit,

a computer processor, and a memory coupled to the computer processor, said memory storing said image, and storing

a) an algorithm for identifying a consecutive sequence of vessel segments until a stop criterion is fulfilled, and an algorithm for each vessel segment determining a vessel state having a vessel state significance, optionally additionally determining a filtered vessel state,

b) an algorithm for selecting the greatest consecutive sequence of vessel segments, wherein each vessel segment has a vessel state significance above a predetermined threshold, and

c) an algorithm for identifying the consecutive sequence of vessel segments se- lected in step b) as a vessel.

Said systems are capable of incorporating any of the variations of the methods described herein.

Legend to Figures

Figure 1 : Fundus image.

Figure 2: Fine Validation.

Figure 3: Search regions for direction measurement.

Figure 4: Principle of direction measurement at the right vessel border.

Figure 5: Examples of directional border response distributions.

Figure 6: The search regions of the center line detector.

Figure 7: Flow diagram for direction measurement. Figure 8: Propagated vessel center point and border search intervals.

Figure 9: Determination of the next vessel center point.

Figure 10: Example of a result of the tracking algorithm.

Figure 11: Example of vessel state significance and two filtered vessel state significances.

Figure 12: Intermediate results of the method of the tracking processes.

Figure 13: (a) Two tracking processes encroaching upon an obstructing tracking process, (b) A failing and a successful pervasion attempt.

Figure 14: Intermediate results of the methods during the engagement phase.

Figure 15: (a) Principle of identifying a suitable vessel state representing the ob- structor. (b) Classification of supplemental region as vessel in order to avoid cleft.

Definitions

Fovea: The term is used in its normal anatomical meaning, i.e. the spot in retina having a great concentration of cones giving rise to the vision. Fovea and the term "macula lutea" are used as synonyms. Also fovea has an increased pigmentation.

Image: The term image is used to describe a representation of the region to be examined, i.e. the term image includes 1 -dimensional representations, 2-dimensional representations, 3-dimensionals representations as well as n-dimensional representatives. Thus, the term image includes a volume of the region, a matrix of the region as well as an array of information of the region.

Optic nerve head: The term is used in its normal anatomical meaning, i.e. the area in the fundus of the eye where the optic nerve enters the retina. Synonyms for the area are, for example, the "blind" spot, the papilla, or the optic disk. Red-green-blue image: The term relates to the image having the red channel, the green channel and the blue channel, also called the RBG image.

Representative for a vessel: The term representative means that the starting point may represent a point or an area of the vessel.

Significance of a vessel or of the optic nerve head: The term describes the distinct- iveness of the vessel or of the optic nerve head.

Significance of a vessel: The term describes preferably the distintiveness of the vessel derived by the distinctiveness of the two individual vessel edges, such as by averaging or by choosing the minimum of the two. The distinctiveness of an individual vessel edge may be represented by the directional derivative or gradient magnitude, and may be computed by pointwise probing or directional averaging.

Starting point: The term describes a point or area for starting the search for candidate optic nerve head areas. The term starting point is thus not limited to a mathematical point, such as not limited to a pixel, but merely denotes a localisation for starting search.

Stop criterion: The term is used to describe a termination threshold for the tracking method, i.e. a stop criterion is any predetermined criterion for terminating the tracking method, such as departure from the image, i.e. the border of the image has been reached, lack of/diminishing significance or interference with already detected ves- sels.

Vessel state: Vessel state is used to describe a vector comprising several elements or vectors relating to the vessel segment determined. For each vessel segment a vessel state is determined.

Vessel state significance: The term describes the distinctiveness of the vessel state, such as the directional derivative of the least distinct edge of the vessel. The significance may also be described as the average directional derivative of the two edges. The distinctiveness of an individual vessel edge may be represented by the direc- tional derivative or gradient magnitude, and may be computed by pointwise probing or directional averaging.

Visibility: The term visibility is used in the normal meaning of the word, i.e. how visi- ble a lesion or a structure of the fundus region is compared to background and other structures/lesions.

Width: The term describes the orthogonal distance between the edges of the vessel. The term is used synonymously with the term caliber.

Detailed description of the invention

Images

The images of the present invention may be any sort of images and presentations of the region of interest. Fundus image is a conventional tool for examining retina and may be recorded on any suitable means. In one embodiment the image is presented on a medium selected from dias, paper photos or digital photos. However, the image may be any other kind of representation, such as a presentation on an array of ele- ments, for example a CCD.

The image may be a grey-toned image or a colour image, in a preferred embodiment the image is a colour image.

Establishing starting points

The starting points may be established by a variety of suitable methods and of combinations of such methods. The variability of fundus images is particularly relevant regarding image dynamics; the contrast may vary considerably from image to image and even from region to region in the same fundus image. A proper tracking algorithm should recognize this circumstance and seek to adapt its sensitivity to the image at hand. Usually, the problem is addressed by initially subjecting the image to some kind of normalization operation. A wide range of possible methods for doing this exist. Local histogram equalization e.g. would be a suitable method for normali- zation of fundus images, but unfortunately this technique is characterized by being rather computationally expensive. The image may be filtered and/or blurred before establishing or as a part of establishing starting points for the method. For example the low frequencies of the image may be removed before establishing starting points. Also, the image may be unsharp filtered, for example by median or mean filtering the image and subtracting the filtered result from the image.

In the interest of computability it is preferred to employ a somewhat simpler method, namely that of subtracting a low pass filtered image from the original and subsequently dividing it by a standard deviation filtered image. The tracking algorithm is ba- sed on adaptive exploratory processing of the normalized image.

In order to enable the tracking algorithm to handle even very small vessels (with a caliber down to about 3 pixels), the tracking algorithm operates in a continuous domain, implying that the local vessel characteristica: Position, direction, caliber etc. be represented by real numbers rather than integers.

Independent of whether the image is filtered or not the starting points may be established as extrema of the image, such as local extrema. Preferably the image is, however, a filtered image, wherein the filtering may be linear and/or non-linear.

In one embodiment the filtering method is a template matching method, wherein the template may exhibit any suitable geometry for identifying the vessels. Examples of template matching are filtering with rectangular filters having various orientations.

It is within the scope of the invention, that the image may be filtered with one or more filters before establishing starting points, or as a part of the step of establishing starting points. Thus, in one embodiment of the invention starting points are established by combining two or more filters.

The extrema may thus be identified indidually by one or more of several methods, such as the following:

The vessels will normally be dark areas in the image, or at least locally the darkest areas. Thus, a method may be establishing at least one intensity extremum in the image, preferably at least one intensity minimum. Therefore, in a preferred embodi- ment, at least one local intensity minimum is established. The extrema may be established on any image function, such as wherein the image function is the unsharped image, the red channel image, the green channel image, or any combinations thereof.

Instead of using intensity or in addition to using intensity the method may include establishing at least one variance extremum in the image, preferably establishing at least one variance maximum in the image. For the same reasons as described with respect to the intensity at least one local variance maximum is established. The ex- trema may be established on any image function, such as wherein the image function is the unsharped image, the red channel image, the green channel image, or any combinations thereof.

In a preferred embodiment the variance extremum is a weighted variance maximum, or a local variance maximum, more preferably a local weighted variance maximum.

Another method for establishing starting points may be random establishment of starting points, wherein the ultimative random establishment is establishing a starting point in substantially each pixel of the image. Of course a random establishment may be combined with any of the methods discussed above.

In yet a further embodiment the starting points may be established as grid points, such as evenly distributed or unevenly distributed grid points. Again this method may be combined with any of the methods of establishing extrema in the image and/or random establishment.

In a preferred method starting points are established by more than one of the methods described in order to increase the propability of assessing the correct localisation of the vessels, also with respect to images having less optimal illumination or presenting other forms of less optimal image quality. A problem that increases when fundus images are recorded decentrally and by less experienced staff than what is the case at specialised hospital departments.

In a more preferred embodiment, the starting points are established by localising the local minima of a mildly blurred version of the original image, and let them act as starting points. Normally a starting point generator is required to produce at least one and preferably more marker points for each vessel segment present in the image.

Validating starting points

Prior to using any of the starting points for tracking, they may be checked for validity using a set of strict validation and verification rules. The purpose of validation is to determine, whether or not the concerned starting point has a great propability for being situated on a blood vessel. If this seems to be the case, the initial local vessel characteristics are estimated for the starting point, and the propagation of these vessel parameters, i.e. the tracking proces may be initiated, as will be described below. Accordingly, the method according to the invention may comprise a step of validating the starting point before identifying characteristics.

The validation may include that the validity of a starting point is represented by a function of the gradient magnitudes at the two borders of the vessel segment or of the gradient at the two borders of the vessel segment.

Thus, in one embodiment, for each of a number of directions distributed evenly around the starting point, the algorithm defines a line piece of a given length and orients it orthogonally to the direction in question. This line piece is placed with its median coinciding with the starting point and then moved outwards until a given maximum distance is reached. For each position of the line piece, the average di- rectional derivative for the line piece in the direction in question is computed. For each direction, the maximum average directional derivative as well as its corresponding offset (distance to the starting point) is stored. The result of this procedure is again a polar function, which must fullfill the following requirements in order for the starting point to avoid being rejected.

a) The maximum sum of function values for two opposite directions must exceed a given threshold. b) Each of the two directions considered in a) must approximately describe a local maximum. A principle of validation is depicted in figure 2.

Characteristics

From the starting point, or optionally validated starting point, at least two characteristics of a vessel segment are detected. By the term characteristics is meant parameters of the vessel segment, i.e. parameters describing the vessel segment. The characteristics may be any parameters describing the vessel segment, and are preferably selected from the significance of the vessel segment, the validity of the starting point, the width of the vessel segment, the direction of the vessel segment, the x- position of the centre point of the vessel, and/or the y- position of the centre point of the vessel, the x- position of a left edge point of the vessel, and/or the y- position of a left edge point of the vessel, and/or the x- position of a right edge point of the vessel, and/or the y- position of a right edge point of the vessel, curvature, colour, and/or intensity.

The present inventors have found that a robust and reliable detection should preferably include the width of the vessel segment and/or the direction of the vessel segment among the characteristics. More preferably the characteristics include the width of the vessel segment and the direction of the vessel segment.

The direction, θ , in which the vessel proceeds can be measured in a variety of ways. One possibility is to seek for the large gradients associated to vessel borders. Another possibility is to utilize the fact that vessels are generally darker than their surroundings and thus seek for the small intensity values associated to the vessel center.

The width is normally detected at the orthogonal distance between the two edges.

At least two characteristics are detected for each vessel segment in order to obtain a robust method, however the more characteristics the better. Therefore, it is preferred that at least three characteristics are determined, such as at least four characteristics. Once the characteristics of a vessel segment are detected, the starting point may be corrected with respect to one or more of the characteristics, in order to perform a corrected starting point for the tracking algorithm for tracking a neighbouring vessel segment.

Vessel states

Furthermore, a vessel state comprising information of at least one of the characteristics detected is determined. Vessel state is used to describe a state comprising one or more elements or vectors relating to the vessel segment determined. For each vessel segment a vessel state is determined. It is furthermore preferred that the significance of the vessel state is determined, wherein the significance of the vessel state describes the distinctiveness of the vessel state, such as the directional derivative of the least distinct edge of the vessel. The significance may also be de- scribed as the average directional derivative of the two edges.

In a preferred embodiment, the vessel state is determined as described in the following:

The orientation lying orthogonally to the two directions considered above is designated as being descriptive of the orientation of the vessel and denoted θ₀. Note that this orientation corresponds to two opposite directions. The sum of the two corresponding offsets is designated as the caliber of the vessel and denoted ω_Q . The point lying directly between the two points representing the two maxima is desig- nated as being descriptive of the center of the vessel and denoted (x_Q, y₀) . In this embodiment this point also constitutes a correction of the starting point. Whereas the starting point is described in integer precision, the four parameters, θ₀ , ω_Q , x₀ , and y₀ are all described in double precision. These four characteristics give an adequate description of the strictly local vessel morphology. However, as will be- come clear below, it is convenient to add a fifth parameter to these, namely a parameter, S₀ representing the local vessel significance and this quintet is referred to as a vessel state in the following. As may be clear from the explanation above, two oppositely directed vessel states may be determined in step b), i.e. one for each of the directions described above, each of said vessel states comprising information of the characteristics detected.

Priority

The tracking may be conducted by at least two principally different methods:

Tracking a vessel or a partial vessel until a stop criterion has been fulfilled, where- after another vessel or partial vessel is tracked, or

Tracking by "jumping" from starting point or corrected starting point in various tracked vessels parts, either at random or preferably based on a priority ranking.

The latter method has the advantage that a starting point having the highest distinc- tivenes continuously gives rise to the tracking, leading to a more robust algorithm, since vessel parts being difficult to tracked from one starting point, may be more securely tracked from another.

Therefore the selection of at least one starting point representative for a vessel in step a) may include selection and optionally validation of all starting points established before continuing with the tracking as well as selection of one starting point, tracking therefrom, and only selecting the next starting point, once the tracking of the former vessel part has finalised. And of course the method may include any combinations thereof, such as selection of a part of the starting points established, tracking therefrom, and then selecting another part of the starting points.

The priority of the starting points may be conducted in any suitable manner; however, it is preferred that the priority is based on the vessel characteristics deter- mined, and more preferably that the priority is a priority of the vessel states. Accordingly, in one embodiment of the invention, at least two vessel states are ranked in a priority queue, such as ranked according to their significance.

The tracking of the vessels in an image may then perform as follows: For all starting points or validated starting points, the related vessel states are ranked in a priority queue, for example according to their significance. The tracking starts with the highest ranking vessel state, after identification of a neighbouring vessel segment and determination of a vessel state for the neightbouring vessel segment, the vessel states are re-ranked in the priority queue with respect to the new vessel state. The tracking continues from the starting point having the highest ranking vessel state, that may be the vessel state of the segment just determined or any other vessel state in the image. Thereby the tracking "jumps" back and forth depending on the ranking and re-ranking of the vessel states. In this regard, only vessel states in the ends of the vessels tracked are ranked in the priority queue, whereas vessel states of vessel segments being positioned between other tracked vessel segments are preferably being stored in a memory, for later retrieval in case the tracked vessel is rejected at a later stage.

Thus, in a preferred embodiment a neighbouring vessel segment is identified in step c) from a starting point related to the highest ranking vessel state in the priority queue.

Weighting - Kalman

Due to the circumstances considered regarding image quality, noise, and variability among fundus images, a tracking algorithm solely made up of independent consecutive measurements of vessel characteristica such as the location of the vessel center or vessel borders, is likely to produce poor, rough or even incorrect results, especially when confronted with low-quality images. A tracking algorithm ought to take into account the possibility of vessel measurements being infected by noise or being downright corrupted by interfering objects such as other vessels or pathologies. Instead of relying completely on the vessel measurements, the tracking algorithm should utilize knowledge regarding the general characteristics of retinal ves- sels of generally not exercising sudden abrubt changes in direction or caliber. Figuratively speaking, the tracking algorithm should assign limited amounts of confidence in measurements suggesting such abrupt changes, and instead assume that the propagated vessel parameters are to some extent in accordance with what could be expected. Another way of putting this is to say that the tracking to some extent

SUBSTITUTE SHEET (RULE 26) should exert skepticism with regard to vessel parameter measurements as it propagates along the vessels.

The optimal linear estimator or Kalman filter constitutes a suitable tool for imple- menting this sort of behaviour. Kalman filtering is a general and powerful methodology for correcting noise-influenced measurements of a given process according to any apriori knowledge available regarding the process.

From a starting point, optionally a validated starting point, the next step is to carry out a tracking on the basis of the initial vessel characteristics. In this example the preferred characteristics have been selected as the width of the vessel, the direction of the vessel, as well and the coordinates of the centre point, i..e (x₀, y₀, θ₀, ω₀), which was estimated as described above. The tracking is a sequence of exploratory searches, in which the vessel parameters (x, y, θ, ω) are propagated along both directions of the concerned vessel. The tracking is terminated, if it encounters a stopping criteria such as lack of significance or departure from the image region or interference with other processes, such as interference with already tracked vessels.

In a preferred embodiment the tracking algorithm is inclined to follow the most prominent vessel border. This improves its ability to breach crossings and bifurcations and in general to handle vessels, where one border for some reason is obscured.

For at least one of the characteristics detected, the characteristic is weighted with an apriori estimate of said characteristic, thereby obtaining a validated characteristic. It is preferred that all characteristics being part of the vessel state are weighted as described.

First, proper assessment of the quality of a priori estimates associated to a given process, requires a sufficiently descriptive mathematical model of the process. An accurate model involving all of the relevant aspects of the given process is required in order to be able to produce a priori estimates of a high quality. However, the establishment of an elaborate mathematical model of the vessels of the human retina is impeded by the extensive diversity displayed by retinal images. Besides, mat- ters are only made worse by the obvious requirement of having the model take into account the various phenomena displayed by the pathological images of most interest. Consequently, a general mathematical model of the retinal vessels may very well prove to be more appropriate than a more specific model.

Secondly, the task of assessing the quality of the measurements of vessel state parameters may at a first glance seem more feasible that what is the case for the corresponding a priori estimates of the vessel state parameters. For instance, if a measurement of a vessel state parameter is characterized by having a small associated variance, it may seem appropriate to interpret this small variance as being indicative of a high quality of the measurement. However, a measurement being specific (having a small variance) does not necessarily imply that it is correct. It may be corrupted by an interferring object, such as another vessel or a pathology.

In consideration of these aspects and the general desirability of establishing a clear and computationally inexpensive tracking algorithm, a very elementary mathematical model of the vessels has been chosen. Furthermore, rather than trying to assess the quality of a priori estimates and measurements of vessel state parameters, the tracking algorithm merely assumes that the relative quality of an a priori estimate and a measurement can essentially be described by a fixed predefined ratio.

Rather than performing an overall measurement of the next vessel state, and subsequently exposing it to a multi-dimensional Kalman correction, it has turned out to be an advantage to split up the estimation into a sequence of individual characteristic measurements and expose each of these to 1 -dimensional Kalman corrections. Apart from generally increasing the performance of the tracking algorithm, this approach makes it easier to adjust the behaviour of the tracking by tuning its associated constants. One of the forces of Kalman filtering and related approaches is the concept of combining the outputs of multiple and possibly different measurements devices in a single measurement of higher quality than each of the individual mea- surements.

The blending factor, K is usually calculated in consideration of the quality of the a priori estimate and measurement, respectively, i.e. the amount of confidence which can be assigned to these two entities. Formally, this approach implies, that the va- lues of the blending factors, K , in the tracking algorithm can be described by con- stants. Accordingly, it is preferred that the detected characteristic is weighted with the estimated characteristic at a predetermined constant ratio.

The constants are suitably determined by taking into account how far from a starting point or a corrected starting point the neighbouring vessel segment is identified, this distance being called the look-ahead-distance. This is another way of saying that the above discussed skepticism with respect to the detected characteristics should increase by decreasing look-ahead-distance, i.e. the shorter from the starting point the neighbouring vessel segment is identified, the more probable it is that only very slight deviations from the former vessel segment are foreseen, whereas a long look- ahead-distance gives rise to a lesser skepticism to the actually detected characteristics.

Thus, for many embodiments the detected width is weighted with the estimated width at a constant ratio, said ratio preferably being in the range of from 0.01 to 1.0, such as in the range of from 0.01 to 0.8, such as in the range of from 0.03 to 0.15, such as in the range of from 0.04 to 0.1 , and the detected direction is weighted with the estimated direction at a constant ratio, said ratio being in the range of from 0.01 to 1.0, such as in the range of from 0.1 to 0.8, such as in the range of from 0.3 to 0.7, such as in the range of from 0.4 to 0.6, wherein a small constant means a great skepticism with respect to the detected characteristics, and a great constant means only a little skepticism with respect to the detected characteristics.

Furthermore, the thus validated coordinates of the centre of the vessel segment are mostly a suitable corrected starting point for the tracking.

Look-ahead-distance

As discussed above the characteristics for the neighbouring vessel segment is pref- erably detected in a predetermined look ahead distance from the, optionally corrected, starting point. Said look-ahead-distance may be determined by in principle two different manners, either as a predetermined constant distance, or more preferred a distance that is dynamically adapted to the image, by calculating the distance from parameters in the image. Accordingly, in a preferred embodiment the look ahead distance towards a neighbouring segment is determined based on the validated characteristics of the vessel segment(s) determined, more preferably based on the validated width of the vessel segment just determined. Thereby the look-ahead-distance decreases with decreasing width of the vessel which is suitable for most purposes. Typically the look ahead distance is determined by multiplying a constant with the validated width of the vessel segment just determined, such as a constant is selected in the range of from 0.01 to 5.0, such as from 0.10 to 2.50, such as from 0.15 to 1.0, such as from 0.2 to 0.8, such as from 0.2 to 0.5.

Tracking

The tracking is an iterative process, segment for segment, until all vessels to be determined are determined.

In the following the tracking is exemplified with respect to vessels in a fundus image based on the preferred characteristics discussed above.

During tracking, the vessel state, ψ is represented by the same set of parameters as the initial parameter set produced by the validation, i.e. ψ =

, where the subscript n denotes an iteration counter. Other representations are pos- sible; a vessel state could for instance be represented by the coordinates of the two vessel border points or by adding further descriptive parameters such as local curvature.

In the following, the method employed by the tracking algorithm for measuring direc- tions is described. Whereas the method does not directly make use of Kalman filters, it is still in possesion of some of the qualities characterizing Kalman methodology.

The backbone of the direction measurement method is 3 measurement devices re- presented by the 3 fan-shaped search regions in figure 3. The innermost measurement device acts as a center line detector, whereas the outermost measurement devices act as border detectors. The estimate of the vessel center point, (x„,y„) , constitutes the central point of the innermost search region. At the initial iteration, (x_n,y„) is set to (x_ϋ,y₀) , i.e. the center point produced during fine validation. The estimates of the corresponding two border points constitute the central points of the two outermost search regions. The search regions are directed along θ^~„+ι , i.e. the a priori estimate of θ_n+ . The direction a priori estimate at the initial iteration, i.e. θ^~ equals θ₀ , the direction produced during fine validation. Besides, the search regions are specified by a search angle, φ , as well as a look ahead distance, D_u , as indicated in figure 2. φ is constant throughout the tracking process, but D_LA is computed for each iteration by multiplying the current caliber estimate, ώ_n of the vessel by a predefined factor, P_u . Hereby, the algorithm adapts to the scale of the vessel being tracked. Scale-adaptivity is a valuable attribute during tracking, since the de- tail of measurements can be adjusted to the given scale, and excessive usage of processing power can be avoided. In the present implementation, φ is set to π , whereas P_u is set to 1.6.

During border detection, a line piece sweeps through the directions of the search angle, φ , as illustrated in figure 4. For each position of this line piece, the algoritm computes the average directional derivative (from now on referred to as a border response) in the direction depicted by the small vector in figure 4, i.e. orthogonally to the line piece and directed outwards relative to the vessel being tracked.

Rather than simply storing the specific directions yielding the maximum border responses, the border detectors store the entire ranges of responses and so to speak regards them as representing directional probability density functions, as illustrated in figure 5. From now on, a range of directional responses representing a directional probability density function, will briefly be referred to as a signal.

While the average directional derivatives are appropriate as border responses, the average image intensities are more appropriate as center line responses. Notice that the line piece corresponding to the direction of the vessel center is characterized by having relatively small average image intensity, so the center line detector should store the negated average image intensities.

However, while this approach performs well on most vessels, it is not well suited in handling the considerable number of vessels displaying axial reflex, where the cen- tral vessel region due to certain imaging conditions appear brighter than the rest of the vessels. This phenomenon is especially marked for the largest arteries in fundus images. In order to counter this problem, the center line detector does preferably not employ the search region depicted in figure 3, but instead makes use of two search regions, arranged with their central points corresponding to the medians of the left and right half of the line piece representing the vessel caliber, as illustrated in figure 6.

For each of these two search regions, a signal is produced by computing the ne- gated average image intensities in the directions of the search angle, as was illustrated for the simple center line detector. A signal produced by one of these search regions is in itself of little use, since it will be biased inwardly for vessels not displaying axial reflex. However, the signals from the two search regions will in this case be oppositely biased, and as such, the sum of the signals will in general con- stitute a proper representation of a directional probability density function, as conceived by the center line detector.

The next step of the direction measurement is to combine the signals produced by the border detectors with the signal produced by the center line detector, thus achi- eving an overall directional probability density function.

Due to the equivalent nature of the two border detectors, the signals produced by them can be added directly, thus yielding an overall border signal. On the other hand, the different nature characterizing the signal produced by the center line de- tector, necessitates a different method of combining it with the overall border signal.

The present algorithm employs a method of normalizing the two signals prior to combining them. Normalization is done by subtracting the minimum occuring value of a signal from all the values of the signal and subsequently divide each signal va- lue with the new sum of signal values. The two resulting signals will then be characterized by each having a minimum value of 0 and a sum of 1. After normalization, the signal associated to the center line detector is multiplied by an arbitrary weight factor, Q , and the two signals are added. In the present implementation, Q is set to ¹/₂. The normalization operation ensures that the contributions from the border detectors and the center line detector to the direction measurement will be suitably balanced, regardless of the brightness and contrast of the fundus image at hand. As an example, this method posseses the quality of being independent of an eventual additive or multiplicative constant being applied to the image function.

The final operation in calculation an overall directional probability density function is the multiplication of the weighted sum of signals by a gaussian distribution, centered at θ^~ _+l , and having a standard deviation of σ . The purpose of this operation is to reduce confidence in large directional responses lying peripherally relative to the a priori estimate of the direction, θ^~ _+l . In the present implementation, σ is set to 1.0 rad.

Ultimately, the direction representing the maximum occurring value in the resulting directional probability density function is designated as θ_{z n+l} , i.e. a measurement of the direction in which the tracking algorithm is to proceed. The procedure of direction measurement is illustrated by the flow diagram in figure 7.

The next step of the tracking algorithm is to subject θ to an actual Kalman filtering operation, combining the direction measurement θ_{z n+l} and the a priori estimate of the direction, θ^~ _+l in an a posteriori estimate of the direction, θ_n+ . In correspondence with the very elementary model of the vessels, it is simply assumed that the next direction will be identical to the current direction. Formally, this implies that the a priori estimate of the next direction, θ^~ _+l is assigned the value of θ_n , i.e. the a poste- riori estimate of the current direction. As stated before, the direction yielded by the fine validation, θ₀ , acts as θ^~ , i.e. the direction a priori estimate at the intial iteration.

Due to the circumstance that θ constitutes a cyclic variable, it would be erroneous to employ the Kalman equation directly. First of all, it makes no sense to calculate the average of two directions separated by an angle of π . In other cases it will be reasonable to define an average as being situated in the smallest interval separating the two directions, but if the phase shift of the direction variable is also situated in this interval, an ordinary averaging operation will produce an incorrect result.

This problem can be overcome by adjusting the two directions prior to calculating their average and subsequently readjust them, in accordance with the following e- quations:

Figuratively speaking, this operation corresponds to calculating the direction avera- ge in a coordinate system, which has been rotated in such a way that the phase shift is placed in the opposite direction of one of the two participating directions, here the direction a priori estimate, θ^~ _+l . As mentioned previously, the direction blending factor, K_g is described by a fixed ratio. In the present implementation K_θ is set to

0.5, implying that the direction apriori estimate and the direction measurement will be assigned equal amounts of confidence.

The a posteriori estimate of the next direction, θ_n+l , provides a basis for the subsequent calculations of the tracking algorithm. Furthermore, 0„₊₁ is final in the sense that it will not be subjected to further corrections until the next iteration of the track- ing process, and as such, θ_n+ is equivalent to the third parameter of the next overall vessel state estimate ψ

Following the calculation of θ_n+ , attention is directed at the other 3 parameters of the vessel state. First, the vessel center point, (x_n,y„) is propagated a certain di- stance, D_p in the direction of θ_n+l , as illustrated in figure 6. As was the case for the look ahead distance, D_u , which was used during direction measurement, the propagation distance, D_p is computed for each iteration by multiplying the current cali- ber estimate, ω_n by a predefined factor, P_p . In the present implementation, P_p is set to * .

Contrary to what was the case for θ , the coordinates obtained by propagating (x„,y„) are at this stage preferably only preliminary; they will preferably be subjected to further corrections before the entire vessel state estimate, ψ will be available. In the following, the preliminarily propagated vessel center point coordinates will be referred to as (x_p,y_p) .

The next step of the tracking algorithm is to search for the locations of the two vessel borders along a line passing through the propagated vessel center point, and lying perpendicular to the estimated vessel direction, θ_n+ , as illustrated in figure 8.

Consideration is restricted to two closed intervals lying symmetrically around the propagated vessel center point. Whether a point of such an interval corresponds to the vessel border or not can be determined by computing the directional derivative of the image function in this point in a direction perpendicular to the direction #_π+1 , and directed away from the vessel, as indicated by the small vector in figure 8. An even better approach is to let the point describe the median of a line piece lying pa- rallel to the direction θ_n+i , and compute the average directional derivative of this line piece, in the same direction as before. Averaging contributes to the suppression of noise, and generally results in a more robust tracking algorithm. In the interest of scale-adaptivity, the width of the border search intervals, D_w , as well as the length of the line pieces used for computing the average directional derivatives, D_L are again calculated by multiplying the current caliber estimate, ώ_n of the vessel by predefined factors; P_w and P_L respectivelely. In the present implementation, P_w is set to 0.6, whereas P_L is set to 1.6.

The locations of the two maximum average directional derivatives of the two inter- vals correspond to a measurement of the locations of the two vessel borders. Still, as was the case for the measurement of direction, the skeptic tracking algorithm does not necessarily accept the measured vessel borders as being in correspondence with the actual vessel borders. In order to correct the vessel border measurement, the algorithm utilizes the circumstance, that the operation of measuring the location of vessel borders implies measuring the caliber of the following vessel state, ω_n+l . As was the case for the vessel direction, the elementary vessel model suggests that the next vessel caliber will be identical to the current caliber. Formally speaking, the a priori estimate of the next caliber, ω^~ _+l is assigned the value of ώ_n , i.e. the a posteriori estimate of the current caliber. As before, the caliber yielded by the fine validation, ω₀ , acts as ω^~ , i.e. the caliber a priori estimate at the intial itera- tion. We can thus calculate an a posteriori estimate of the caliber of the next vessel state, ώ_π+1 , by the equation:

ω»₊ι = <»;₊, + ^κω ^• ,„₊ι ^" ok. )

Here, ω_{2 π+1} denotes the measurement of the caliber of the next vessel state, i.e. the distance between the two points describing the measured locations of the two vessel borders. K_ω denotes the caliber blending factor, which is again described by a fixed ratio. In the present implementation, K_ω is set to 0.05. Contrary to what was the case for the direction blending factor, K_θ , the small value of K_ω implies a pro- foundly asymmetrical assignment of confidence between the caliber a priori estimate, ω^~ _+\ and the caliber measurement. The weight of ω^~ _+l is 20 times the weight of ω_{z n+1} , and as such, the algorithm is very reluctant in accepting a caliber measurement which is not in correspondence with the preceeding caliber measurement.

The a posteriori estimate of the vessel caliber, ώ_n+l , is by now final; it will not be subjected to further corrections until the next iteration and thus corresponds to the fourth parameter of the next overall vessel state estimate ψ

The single remaining question in the process of estimating ψ is to estimate the coordinates of the next vessel center point, (x_n+l,y„_+l) ■ One simple possibility would be to keep the point (x_p,y_p) , i.e. the coordinates obtained by propagating the center point along the a posteriori estimate of the direction θ_π+l , as was illustrated in figure 8. However, the present tracking algorithm employs a more elaborate method, which is in better correspondence with our intent of following the most prominent border. This method is described in the following.

The problem of estimating the coordinates of the next vessel center point, (x_π+l,y_n+l) can geometrically be interpreted as a matter of placing the line piece representing the next vessel state suitably relative to the preliminarily propagated vessel center point (x_p,y_p) . This situation is illustrated in figure 9, where the concerned line piece is drawn in bold.

Some of the characteristics of this line piece have already been determined: it is oriented orthogonally to θ_n+i , and its length is given by ώ_n+1 . Furthermore, it is rea- sonable to employ the restriction that the line piece must pass through (x_p,y_p) . This leaves the line piece with only one degree of freedom, namely the distance from the median of the line piece to (x_p,y_p) , denoted by s in figure 9. If the median of the line piece is situated to the left of (x_p,y_p) with respect to θ_n+l , s is indicated as being negative, and vice versa if the median is situated to the right of (x_p,y_p) . This signed distance, s , will be referred to as a shift from now on.

The problem of estimating (x_π+l,y_n+l) thus becomes a problem of estimating the shift parameter, s_n+l . One obvious choice might be to place the line piece centrally relative to the two measured border points, represented by the small dots in figure 9. This would correspond to a shift parameter of s = ' > • (AR - AL) , where ΔZ, and ΔR describe the distance from (x_p,y_p) to the measured left and right border points, respectively. However, in doing so the initially stated intention of following the most prominent vessel border would be neglected. In the present tracking algorithm, the magnitudes of the two maximum average directional derivatives or border respon- ses, which were calculated during border detection, are regarded as being indicative of the prominence of the respective vessel borders. Based on these two border re- sponses the algorithm then computes what we have chosen to term a priority fraction, F according to the following equation:

_ _ max(0,E_Λ)^α max(0, E_L )^α + max(0, E_R )^α

Here, E_L and E_R denotes the maximum border response for the left and right border respectively, α is a predefined exponent with the purpose of strengthening the confidence in the most prominent border at the expense of confidence in the least prominent border. In the present implementation, α is set to 7. Notice that such a large exponent implies that even a small difference in the two border responses will have a substantial influence on the value of F . The purpose of the three max operators in the equation is to ensure that an eventual negative maximum border response will not corrupt the calculations.

Thus, the resulting value of F will be situated in the interval [0; l] . A value of F = 0 implies full confidence in the left border measurement and no confidence in the right border measurement, and vice versa if F = 1. In the first case the line piece representing the next vessel state is placed in such a way that its left end point coincides with the measured left border point. In the latter case, the line piece is placed in such a way that its right end point coincides with the measured right border point. The placement of the line piece for intermediate values of F varies linearly with F . The following equation describes the formal implications of this method regarding the shift parameter, s .

5_2ι„₊₁ = '/₂ - (ΔR - Δ ) + (E - '/2) - „₊₁ - ώ„₊₁)

The calculated shift parameter represents a measurement, emphasized by the subscript z in the equation above. As was the case for the measurement of direction and caliber for the next vessel state, the shift measurement is not necessarily trusted by the skeptic tracking algorithm, and consequently the shift parameter is subjected to a Kalman correction operation, according to the following equation: ^Sn+\ ~ ^Sn+\ ⁺ K_S ^Sz, n+\ ~ ^Sn+l )

In accordance with the elemental vessel model, the shift apriori estimate, s^~ _+l is set to 0, so the equation above can be simplified to:

n+1 = κ. 'z, n+\

In the present implementation, the shift blending factor, K_s is set to 0.4, implying that the contribution of the a priori estimate, s^~ _+l to the shift estimation is slightly larger than the contribution from the measurement, s_z ,.

Having calculated the a posteriori estimate of the shift parameter, _?_π+1 we can perform a final correction of the coordinates of the vessel center point, which will then be in correspondence with the first and second parameters of the next overall vessel state estimate, ψ , respectively:

^cos(0_n+1 - Viπ)

This concludes the current iteration of the tracking algorithm; all 4 parameters of the next overall vessel state, _ ψn+l have been estimated, and provided that the — ψn+l does not meet one of the stopping criteria described below, the tracking can proceed in the same manner as before, i.e. by first considering the directional parameter, θ and so on.

Figure 10 depicts the result of applying the described tracking algorithm to a larger part of the vessel which was also depicted in the other figures 2-8 herein.

The tracking should stop when a stop criterion has been fulfilled. A stop criterion refers to any situation, wherein the tracking should stop. In the present context a stop criterion is preferably at least one of the following criteria: Departure of image,

Interference with other processes, such as interference with already tracked vessels, and

Lack of and/or diminishing significance of the vessel state.

Once a stop criterion has been fulfilled, that tracking method stops, and the vessel part consisting of a consecutive sequence of tracked vessel segments may then be accepted as a vessel.

In a preferred embodiment, the tracking propagates in two directions from the initial starting point, and only after fulfilling a stop criterion for both ends, the vessel part therein between may be accepted as a vessel.

In a preferred embodiment, information regarding vessel segments identified is stored in a memory. Thereby, it is possible to conduct a validation of the vessel part tracked before accepting the tracked vessel as a vessel of the image. Such information is typically related to the significance of the vessel segment, and in particular of the vessel state significance.

By use of this information it may be possible to determine a filtered vessel state significance, wherein the filtering procedure is based on the tracking history of all the other vessel segments tracked, and in particular of the vessel state significance of the other vessel segments. The filtering method allows a vessel segment, the significance of which is below a certain threshold to be accepted, because the neighbouring vessel segments have a higher significance. Thereby vessel segments that would otherwise have been rejected may be accepted due to the validity of their neighbours.

In one embodiment, vessel segments having a vessel state significance and/or a filtered vessel state significance above a predetermined threshold (termination threshold) may be accepted as validated vessel segments of the image. Preferably the vessels tracked essentially only consist of validated vessel segments. Thereby, a stop criterion may be fulfilled when the filtered significance of the vessel state is below a predetermined exclusion or termination threshold. The significance may be filtered from both ends of the vessel part, whereby the vessel segment may be accepted if only one of the filtering methods raises the filtered significance above the threshold. Vide in this respect Fig. 11.

Once a stop criterion has been fulfilled, preferably for both directions of the tracking, the vessel part in between is examined with respect to the validity of the vessel segments therein. The longest vessel part between the two stops comprising a consecutive sequence of validated vessel segments is selected as a vessel part. In case the tracked vessel comprises non-validated vessel segments, these segments are cleared and may thus be part of the next tracking procedure as any other parts of the image. The longest vessel part between the two stops should preferably have at least a predetermined length, i.e. a predetermined number of vessel segments in a consecutive sequence in order to be accepted as a vessel part. Accordingly, the vessel part should preferably comprise at least a predetermined number of vessel segments, wherein the predetermined number is in the range of from 10 to 100, such as in the range of from 25 to 75, such as in the range of from 30 to 50.

Another way of determining the number of vessel segments is that the predeter- mined number is determined as the ratio of the length of the total vessel segments to the average width of the segments.

All the vessel segments participating in the vessel part should preferably have a filtered vessel state significance above a predetermined threshold.

Figure 12 depicts a sequence of intermediate results of the described method of the tracking processes, when applied to a region of an actual fundus image. In the final image of this sequence, all the tracking processes are either suppressed (prevented from initializing), obstructed by other tracking processes, have met the stopping cri- terion related to lack of filtered vessel state significance, or have departed from the image. A total of 1078 iterations were carried out by the algorithm in this example.

The methods described above, where the most promising tracking process is continually being favoured at the expense of the more questionable tracking processes, are in possession of a number of qualities. First, the problem of having the inertness of the tracking processes result in poor delineation of vessels, has to a high degree been disposed of, since the vessel states with the largest distinctness values will in general also be the ones corresponding the most to the actual vessels. So, when a tracking process starts to lose track of a vessel, chances are that control is transfer- red to a better qualified tracking process which manages to catch up and merge with the tracking process which was about to lose track. Furthermore, the possibility of having two tracking processes cooperate by merging into a single tracking process may also bring about the detection of vessels, which would not be detected by the tracking processes operating on an individual basis. As an example of this, consider the situation where a considerable part of a vessel is obscured by some interfering feature. Whereas a tracking process might lack the ability to breach such a region on its own, an encounter with a matching tracking process inside the region might still bring about the detection of the vessel.

When having tracked all vessel segments until a stop criterion is met, it may be advantageous to engage the tracking process in order to ensure that the stop criterion in itself does not lead to undetected vessels. Thus, the engagement method includes that the event of encountering another tracking process constituting a decisive stopping criterion, is abandoned. Instead, a tracking process is allowed to encroach somewhat upon an obstructing tracking process in an attempt to pervade it, as illustrated in figure 13.

Figure 13b illustrates the principle of pervading an obstructing tracking process by allowing the tracking to carry on for a certain limited stretch, thereby identifying ten- tative vessel segments. Tentative vessel segments are shown with dotted lines in fig. 13. If the encroaching tracking process is still obstructed by the other tracking process after having spent the allowed stretch, the pervasion attempt is regarded as having failed, and accordingly the process should be terminated in the given direction. Furthermore, the tracking process should not be allowed to claim the region re- quested by it during encroachment, since it would thus be in dispute with the obstructing tracking process. On the other hand, if the encroaching tracking process manages to pervade the obstructing tracking process prior to having spent the allowed stretch, the attempt is regarded as being successful, and the tracking process should generally be allowed to proceed as if it had not encountered the other track- ing process at all. Moreover, the two tracking processes may in this case generally be regarded as not being in dispute, and accordingly they should both be allowed to claim the region shared by them. In figure 13a the dashed outlines mark the regions requested by two tracking processes at a given point in their respective attempts at pervading an obstructing tracking process. The upper tracking process fails and is consequently denied the region requested by it during encroachment, illustrated by the dashed outline in figure 13b. The lower tracking process, however, succeeds at pervading the obstructing tracking process and is consequently granted the entire region requested by it.

The engagement process may be applied to the tracking process once all vessels have met a stop criterion, or applied during the tracking process each time the tracking process encounters a stop criterion. Independent of the time of applying the engagement process, it may be conducted as described in the following:

Specifically, for each of its two directions, a tracking process maintains a score indicating the maximum number of times it successively has encountered the same different tracking process. A propagated vessel state is said to have experienced an encounter with another tracking process in case the line piece corresponding to the vessel state intersects the region claimed by the other tracking process. The outer- most vessel state with respect to a given direction (i.e. one of the two vessel states representing the given tracking process in the priority queue) is classified as being hampered in case its associated score exceeds 0. Correspondingly, the score associated to a given vessel state will be referred to as the hampered count, N_H , of the vessel state. The stretch a tracking process is allowed to encroach upon an obstruc- ting tracking process is described by a predetermined number of iterations, N_{H Max} , and accordingly, the elaborated stopping criterion may formally be described by the equation N_H > N_{H Max} . In the present context N_{H Max} is preferably at least 5, such as at least 10.

Regarding the incorporation of the modified stopping criterion in the scheduling scheme, the algorithm employs the same principle as before, where the vessel states representing the tracking processes in the priority queue are sorted according to their distinctness. However, the algorithm adds to this the restriction that a hampered vessel state will always have a lower priority than a vessel state which is not hampered. This implies that the algorithm will reach the same condition as before, i.e. where all tracking processes are either obstructed, have departed the image or have met the stopping criterion related to lack of filtered vessel state significance. However, the process may then proceed by propagating the vessel states until they have all either met the modified stopping criterion of having their count exceed N_{H ax} or have met one of the usual stopping criteria.

Accordingly, the present invention relates to a method for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

b) continuing identifying tentative vessel segments from the stop criteria point for a predetermined number of iterations and/or for a predetermined length, and for e- ach of said tentative vessel segments identifying a vessel state having a vessel state significance, optionally additionally determining a filtered vessel state,

c) determining whether the stop criterion is fulfilled for all tentative vessel segments,

d) continuing identification of vessel segments as defined in step a) if at least one tentative vessel segment does not fulfil the stop criteria.

The method may of course be combined with the other methods described herein as discussed above.

Preferably the continued identification in step d) continues from the latest indentified tentative vessel segment.

Figure 14 depicts a sequence of intermediate results of the algorithm during the engagement phase. The black regions are claimed by tracking processes, whereas the dark grey regions are merely requested. As seen in figure 14, the algorithm has managed to correctly identify 3 of the 4 major vessel crossings present in the image. Furthermore, by interpreting the tracking processes ultimately being obstructed by others as representing vessel bifurcations, the algorithm will also have managed to correctly identify 7 vessel bifurcations present in the image.

Accordingly, by the engagement process, vessel crossovers and bifurcations may be identified. When having identified a number of vessel segments as discussed above, it may be preferred to validate the tracking process as such, in particular with respect to a consecutive vessel sequence being so short that it fails to meet a length validation requirement described above.

The present invention also provides a method for validating short consecutive vessel sequences by their relation to other vessels in the image. This validation process includes assigning a probability factor to the vessel sequence depending on the engagement with other identified vessel sequences in one end or both ends, the later having a higher probability factor, leaving short vessel sequences to no validation if they are isolated from the other vessels identified. This is exemplified below in relation to figure 15a and 15b.

Formally, the distance from the obstructed vessel state, Ψ_A to a given vessel state, Ψ_fl of the obstructing tracking process will be given by the equations:

^_A = (θ_A , ω_{A 1} x_Λ , y_A , D_Λ ), ^x¥_B = (θ_B , ω_B,x_B , y_B, D_B)

Q is a weighting factor for the direction parallel to Θ_A relatively to the direction perpendicular to Θ_A . As such, Q corresponds to the ratio of the magnitude of the major and minor half-axis of the ellipsis in figure 15a. In the present implementation, Q has been set to 3.0. If the obstructed tracking process as well as the vessel state representing its obstructor is ultimately accepted, it will be reasonable to classify a supplemental region as vessel, such as illustrated in figure 15b. This is in order to avoid the clefts which would otherwise blemish many of the bifurcations of the final outcome of the algorithm.

Applications

In the following, examples of various applications of the method according to the invention are discussed.

Registration of fundus images

When recording fundus images, at least 4 images are normally recorded from each fundus, representing different regions of the fundus. A golden standard for fundus images is recordation of 7 regions, all overlapping partly at least one of the other. In order to examine the fundus region properly, registering or mounting of the images in a continuous manner with respect to the structures in the image, such as for example by arranging the images so that the vessels correctly continue in the images.

Accordingly, the invention also relates to a method for registering at least two different fundus images of the same fundus, comprising detecting the vessels in said images by a method as defined above, and orienting the images with respect to the vessels.

Arterioles/venoles detection

Having identified the blood vessels in the image, it is often desirable to be able to distinguish between veins and arteries among the blood vessels. This can be important, for example in the diagnosis of venous beading and focal arteriolar narrow- ing.

The vascular system observed in the ocular fundus images is by nature a 2- dimensional projection of a 3-dimensional structure. It is quite difficult in principle to distinguish veins from arteries, solely by looking at isolated vessel segments. How- ever, it has been discovered that effective separation can be achieved by making use of the fact that, individually, the artery structure and the vein vessel structures is each a perfect tree, (i. e., there is one unique path along the vessels from the heart to each capillary and back).

On the retina, the artery and vein structures are each surface filling, so that all tissue is either supplied or drained by specific arteries or veins, respectively.

A method for distinguishing veins from arteries is described in WO 00/65982 (Tor- sana Diabetes Diagnostic A/S), which is based on the realisation that crossings of vessel segments are, for practical purposes, always between a vein and an artery (i.e., crossings between arteries and arteries or between veins and veins are, for practical purposes, non-existent).

Iterative detection of vessels and optic nerve head

In some of the embodiments according to the invention, it is preferred that the estimation of vessels is adjusted with respect to candidate optic nerve head areas appearing in the image. By adjusted is meant either that an iterative estimation of optic nerve head and vessels is conducted, wherein for each iteration, the significance of the localisation of both increases towards a maximum, or that knowledge of the anatomical localisation of vessels adjacent to the optic nerve head is used for locating and/or validating the position of the optic nerve head. For many of these embodiments it is even more preferred that the estimation of candidate optic nerve head areas is preceded by detection of vessels in the image.

Detection of lesions

A very important aspect of the invention is the detection of any lesions of the fundus adjusted with respect to the vessels. Lesions of the retina normally embrace micro- aneurysms and exudates, which show up on fundus images as generally "dot shaped" (i.e. substantially circular) areas. It is of interest to distinguish between such microaneurysms and exudates, and further to distinguish them from other pathologies in the image, such as "cotton wool spots" and hemorrhages. If the optic nerve head area is present in the image, it may give rise to errors when detecting lesions in the image. The lesions may be detected by any suitable method known to the person skilled in the art. A preferred method is described in a co-pending PCT application entitled "Lesion detection in fundus images" by RETINAL YZE A/S.

System

In another aspect the invention further relates to a system for detecting the vessels in a fundus image. Thus, the system according to the invention may be any system capable of conducting the method as described above as well as any combinations thereof within the scope of the invention. The system may thus be defined as a system for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

an image acquisition unit,

a computer processor, and

b) an algorithm for selecting at least one starting point representative for a vessel, detecting at least two characteristics of a vessel segment comprising the starting point, optionally determining a corrected starting point related to the vessel segment, and determining at least one vessel state, said vessel state comprising information of at least one of the characteristics detected,

c) an algorithm for, from an, optionally corrected, starting point, identifying a neighbouring vessel segment by the steps of:

j) an algorithm for optionally repeating steps b) to i) until all starting points from step a) have been selected.

The image acquisition unit may include any suitable apparatus for image acquisition, such as a camera, a digital camera, a CCD array.

Any of the algorithms of the systems described above may be adapted to the various variations of the methods described above. Accordingly, the system may in- elude algorithms to perform any of the methods described above.

A graphical user interface module may operate in conjunction with a display screen of a display monitor. The graphical user interface may be implemented as part of the processing system to receive input data and commands from a conventional key- board and mouse through an interface and display results on a display monitor. For simplicity of the explanation, many components of a conventional computer system have not been discussed such as address buffers, memory buffers, and other standard control circuits because these elements are well known in the art and a detailed description thereof is not necessary for understanding the present invention. Pre-acquired image data can be fed directly into the processing system through a network interface and stored locally on a mass storage device and/or in a memory. Furthermore, image data may also be supplied over a network, through a portable mass storage medium such as a removable hard disk, optical disks, tape drives, or any other type of data transfer and/or storage devices which are known in the art.

One skilled in the art will recognize that a parallel computer platform having multiple processors is also a suitable hardware platform for use with a system according to the present invention. Such a configuration may include, but not be limited to, parallel machines and workstations with multiple processors. The processing system can be a single computer, or several computers can be connected through a communications network to create a logical processing system.

The present system allows the grader, that is the person normally grading the images, to identify the vessels more rapidly and securely, and thereby locate other structures in the image. Also, the present system allows an automatic detection of lesions and other pathologies of the retina without interference from the vessels, again as an aiding tool for the traditional grader.

By use of the present system it is also possible to arrange for recordation of the images at one location and examining them at another location. For example the images may be recorded by any optician or physician or elsewhere and be transported to the examining specialist, either as photos or the like or on digital media. Accord- ingly, by use of the present system the need for decentral centers for recording the image, while maintaining fewer expert graders could be realised.

Furthermore, in addition to the communication of images and medical information between persons involved in the procedure, the network may carry data signals in- eluding control or image adjustment signals by which the expert examining the images at the examining unit directly controls the image acquisition occurring at the recordation localisation, i.e. the acquisition unit. In particular, such command signals as zoom magnification, steering adjustments, and wavelength of field illumination may be selectively varied remotely to achieve desired imaging effect. Thus, ques- tionable tissue structures requiring greater magnification or a different perspective for their elucidation may be quickly resolved without ambiguity by varying such control parameters. Furthermore, by switching illumination wavelengths views may be selectively taken to represent different layers of tissue, or to accentuate imaging of the vasculature and blood flow characteristics. In addition, where a specialized study such as fluorescence imaging is undertaken, the control signals may include time varying signals to initiate stimulation with certain wavelengths of light, to initiate imaging at certain times after stimulation or delivery of dye or drugs, or other such precisely controlled imaging protocols. The digital data signals for these operations may be interfaced to the ophthalmic equipment in a relatively straightforward fashion, provided such equipment already has initiating switches or internal digital circuitry for controlling the particular parameters involved, or is capable of readily adapting electric controls to such control parameters as system focus, illumination and the like.

Also, the examining expert could be able to exert some treatment in the same remote manner. It will be understood that the imaging and ophthalmic treatment instrumentation in this case will generally include a steering and stabilization system which maintains both instruments in alignment and stabilized on the structures appearing in the field of view. However, in view of the small but non-negligible time delays still involved between image acquisition and initiation of diagnostic or treatment activity at the examination site, in this aspect of the invention, the invention contemplates that the system control further includes image identification and correlation software which allows the ophthalmologist at site to identify particular positions in the retinal field of view, such as pinpointing particular vessels or tissue structures, and the image acquisition computer includes image recognition software which enables it to identify patterns in the video frames and correlate the identified position with each image frame as it is acquired at the acquisition site. For example, the image recognition software may lock onto a pattern of retinal vessels. Thus, despite the presence of saccades and other abrupt eye movements of the small retinal field which may occur over relatively brief time intervals, the ophthalmic instrumentation is aimed at the identified site in the field of view and remote treatment is achieved.

In addition to the foregoing operation, the invention further contemplates that the images provided by acquisition unit are processed for photogrammetric analysis of tissue features and optionally blood flow characteristics. This may be accomplished as follows. An image acquired at the recordation unit is sent to an examination unit, where it is displayed on the screen. As indicated schematically in the figure, such an image may include a network of blood vessels having various diameters and lengths. These vessels include both arterial and venous capillaries constituting the blood supply and return network. At the examination unit, the workstation may be equipped with a photogrammetric measurement programme which may for example enable the technician to place a cursor on an imaged vessel, and moving the cursor along the vessel while clicking, have the software automatically determine the width of the vessel and the subvessels to which it is connected, as well as the coordinates thereof.

The software for noting coordinates from the pixel positions and linking displayed features in a record, as well as submodules which determine vessel capacities and the like, are straightforward and readily built up from photogrammetric program techniques. Work station protocols may also be implemented to automatically map the vasculature as described above, or to compare two images taken at historically different times and identify or annotate the changes which have occurred, highlighting for the operator features such as vessel erosion, tissue which has changed col- our, or other differences. In addition, a user graphical interface allows the specialist to type in diagnostic indications linked to the image, or to a particular feature appearing at a location in the image, so that the image or processed version of it becomes more useful.

With suitable training, the relative health of the vessel, its blood carrying capacity and the like may also be visually observed and noted. This photogrammetric analysis allows a road map of the vasculature and its capacity to be compiled, together with annotations as to the extent of tissue health or disease apparent upon such inspection. Thus, a very precise and well-annotated medical record may be readily compiled and may be compared to a previously taken view for detailed evidence of changes over a period of time, or may be compared, for example, to immediately preceding angiographic views in order to assess the actual degree of blood flow occurring therein. As with the ophthalmologist's note pad entries at examination unit, the measurement entries at examination unit become an annotated image record and are stored in the central library as part of the patient's record. Unlike a simple medical record system, the present invention changes the dynamics of patient access to care, and the efficiency of delivery of ophthalmic expertise in a manner that solves an enormous current health care dilemma, namely, the obstacle to proper universal screening for diabetic retinopathy. A basic embodiment of the invention being thus disclosed and described, further variations and modifications will occur to those skilled in the art, and all such variations and modifications are encompassed within the scope of the invention as defined in the claims appended hereto.

References

[1] A. Can, H. Shen, J. N. Turner, H. L. Tanenbaum, B. Roysam, "Rapid Automated Tracing and Feature Extraction from Retinal Fundus Images Using Direct Explo- ratory Algoritms", IEEE Transactions on Information Technology in Biomedicine,

Vol. 3, No. 2, June 1999.

[2] Peter S. Maybeck, "Stochastic Models, Estimation and Control volume 1", Mathematics in Science and Engineering vol. 141 , Academic Press, 1979.

[3] G. Welch, G. Bishop, "An Introduction to the Kalman Filter", Department of Computer Science, University of North Carolina at Chapel Hill, NC 27599-3175.

[4] P. J. Hargrave, "A Tutorial Introduction to Kalman Filtering", STC Technology Ltd., London Road, Harlow, Essex, CM17 9NA.

Claims

1. A method for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

d) detecting corresponding characteristics for the neighbouring vessel segment,

i) repeating steps c) to h) until a stop criterion has been fulfilled, thereby tracking a vessel comprising the vessel segments identified, j) optionally repeating steps b) to i) until all starting points from step a) have been selected.

2. The method of claim 1 , wherein a network of vessels is identified.

3. The method according to any of the preceding claims, wherein the image is an image of the fundus of the eye.

4. The method according to any of the preceding claims, wherein the image is pre- sented on a medium selected from dias, paper photos or digital photos.

5. The method according to any of the preceding claims, wherein the step b) further comprises a step of validating the starting point before identifying characteristics.

6. The method according to claim 5, wherein the validity of a starting point is represented by a function of the gradient magnitudes at the two borders of the vessel segment or of the gradient at the two borders of the vessel segment.

7. The method according to any of the preceding claims, wherein the starting points are established in local extrema of the image.

8. The method according to any of the preceding claims 1-6, wherein the starting points are established by filtering the image.

9. The method according to any of the preceding claims 1-6, wherein the starting points are established by using template matching.

10. The method according to any of the preceding claims, wherein two oppositely directed vessel states are determined in step b), each comprising information of the characteristics detected.

11. The method according to any of the preceding claims, wherein a significance of the vessel state is determined.

12. The method according to any of the preceding claims, wherein the stop criteria is selected from departure from image, lack of significance of vessel state and/or interference with already tracked vessels.

13. The method according to claim 11 or 12, wherein the significance is the directional derivative of the least distinct edge of the vessel segment.

14. The method according to any of the preceding claims, wherein a vessel segment having a vessel state significance above a predetermined threshold is accepted as a validated vessel segment.

15. The method according to claim 14, wherein vessels tracked essentially only consist of validated vessel segments.

16. The method according to any of the preceding claims, wherein the characteristics are selected from the significance of the vessel segment, the validity of the starting point, the width of the vessel segment, the direction of the vessel segment, the x- position of the centre point of the vessel, and/or the y- position of the centre point of the vessel, the x- position of a left edge point of the vessel, and/or the y- position of a left edge point of the vessel, and/or the x- position of a right edge point of the vessel, and/or the y- position of a right edge point of the vessel, curvature, colour, and/or intensity.

17. The method according to any of the preceding claims, wherein the characteris- tics include the width of the vessel segment.

18. The method according to any of the preceding claims, wherein the characteristics include the direction of the vessel segment.

19. The method according to any of the preceding claims, wherein the characteristics include the width of the vessel segment and the direction of the vessel segment.

20. The method according to any of the preceding claims, wherein at least three characteristics are determined, such as at least four characteristics.

21. The method according to any of the preceding claims, wherein at least two vessel states are ranked in a priority queue.

22. The method according to claim 21 , wherein the vessel states are ranked according to their significance.

23. The method according to claim 19, wherein a neighbouring vessel segment is identified in step c) from a starting point related to the highest ranking vessel state.

24. The method according to any of the preceding claims, wherein the detected characteristic is weighted with the estimated characteristic at a predetermined constant ratio.

25. The method according to claim 21 , wherein the detected width is weighted with the estimated width at a constant ratio, said ratio being in the range of from 0.01 to 1.0, such as in the range of from 0.01 to 0.8, such as in the range of from 0.03 to 0.15, such as in the range of from 0.04 to 0.1.

26. The method according to claim 21 , wherein the detected direction is weighted with the estimated direction at a constant ratio, said ratio being in the range of from 0.01 to 1.0, such as in the range of from 0.1 to 0.8, such as in the range of from 0.3 to 0.7, such as in the range of from 0.4 to 0.6.

27. The method according to any of the preceding claims, wherein the corrected starting point is the validated centre of the vessel segment.

28. The method according to any of the preceding claims, wherein corresponding characteristics for the neighbouring vessel segment is detected in a predetermined look ahead distance from the, optionally corrected, starting point.

29. The method according to claim 28, wherein the look ahead distance towards a neighbouring segment is a predetermined constant.

30. The method according to claim 28, wherein the look ahead distance towards a neighbouring segment is determined based on the validated characteristics of the vessel segment(s) determined.

31. The method according to claim 29, wherein the look ahead distance is determined based on the validated width of the vessel segment just determined.

32. The method according to claim 31 , wherein the look ahead distance is determined by multiplying a constant with the validated width of the vessel segment just determined.

33. The method according to claim 32, wherein the constant is selected in the range of from 0.01 to 5.0, such as from 0.10 to 2.50, such as from 0.15 to 1.0, such as from 0.2 to 0.8, such as from 0.2 to 0.5.

34. The method according to any of the preceding claims, wherein the tracking of a vessel is conducted simultaneously in two opposite directions from the initial starting point of step b).

35. The method according to any of the preceding claims, wherein information regarding vessel segments identified is stored in a memory.

36. The method according to claim 35, wherein the information comprises the significance of the vessel state of each vessel segment tracked.

37. The method according to any of the preceding claims, wherein a filtered vessel state significance is determined for each vessel segment.

38. The method according to claim 37, wherein a vessel state is filtered by weighting the vessel state with information regarding previously identified vessel segments in a consecutive sequence of vessel segments.

39. The method according to any of the preceding claims 37-38, wherein a vessel segment having a vessel state significance and/or a filtered vessel state signifi- cance above a predetermined threshold is accepted as a validated vessel segment.

40. The method according to claim 39, wherein vessels tracked essentially only consist of validated vessel segments.

41. The method according to any of the preceding claims 37-40, wherein a stop criterion is fulfilled when the filtered significance of the vessel state is below a predetermined exclusion threshold.

42. The method according to any of the preceding claims, wherein a vessel tracked is accepted as a vessel in the image if it comprises at least a predetermined number of vessel segments in a consecutive sequence.

43. The method according to claim 42, wherein the predetermined number is in the range of from 10 to 100, such as in the range of from 25 to 75, such as in the range of from 30 to 50.

44. The method according to claim 43, wherein the predetermined number is deter- mined as the ratio of the length of the total vessel segments to the average width of the segments.

45. The method according to any of the preceding claims 42-44, wherein the consecutive sequence of vessel segments consists of vessel segments having a filtered vessel state significance above a predetermined threshold.

46. A system for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

an image acquisition unit,

a computer processor, and

a memory coupled to the computer processor, said memory storing said image, and storing a) an algorithm for establishing at least one starting point representative for a vessel in the image,

c) an algorithm for, from a, optionally corrected, starting point, identifying a neighbouring vessel segment by the steps of:

d) detecting corresponding characteristics for the neighbouring vessel segment,

47. A method for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

b) selecting a consecutive sequence of vessel segments, wherein each vessel segment has a vessel state significance and/or a filtered vessel state signifi- cance above a predetermined threshold,

c) identifying the consecutive sequence of vessel segments selected in step b) as a vessel.

48. The method according to claim 47, wherein step a) includes

a1) establishing at least one starting point representative for a vessel in the image,

a2) selecting at least one starting point representative for a vessel, detecting at least two characteristics of a vessel segment comprising the starting point, optionally determining a corrected starting point related to the vessel segment, and determining at least one vessel state, said vessel state comprising information of at least one of the characteristics detected,

a3) from a, optionally corrected, starting point identifying a neighbouring vessel segment by the steps of:

a4) detecting corresponding characteristics for the neighbouring vessel seg- ment,

a5) providing estimates for the characteristics for the neighbouring vessel segment, and a6) for at least one characteristic individually weighting the detected characteristic with the estimated characteristic thereby obtaining a validated characteristic,

a7) determining a vessel state for the neighbouring vessel segment, said vessel state comprising information of at least one of the detected characteristics and/or the validated characteristics,

a8) optionally determining a corrected starting point related to the neighbouring vessel segment,

a9) repeating steps a3) to a8) until a stop criterion has been fulfilled, thereby identifying a consecutive sequence of the vessel segments identified,

a10) optionally repeating steps a2) to a9) until all starting points from step a1) have been selected.

49. The method of claim 47 or 48, wherein a network of vessels is identified.

50. The method according to any of the preceding claims 47-49, wherein the image is an image of the fundus of the eye.

51. The method according to any of the preceding claims 47-50, wherein the image is presented on a medium selected from dias, paper photos or digital photos.

52. The method according to any of the preceding claims 48-51 , wherein the step a2) further comprises a step of validating the starting point before identifying characteristics.

53. The method according to claim 48, wherein the validity of a starting point is rep- resented by a function of the gradient magnitudes at the two borders of the vessel segment or of the gradient at the two borders of the vessel segment.

54. The method according to any of the preceding claims 48-53, wherein the starting points are established in local extrema of the image.

55. The method according to any of the preceding claims 48-53, wherein the starting points are established by filtering the image.

56. The method according to any of the preceding claims 48-53, wherein the starting points are established by using template matching.

57. The method according to any of the preceding claims 47-56, wherein two opposite directed vessel states are determined in step a2), each comprising information of the characteristics detected.

58. The method according to any of the preceding claims 47-57, wherein the stop criteria is selected from end of image, and/or meeting already tracked vessels.

59. The method according to any of the preceding claims 47-58, wherein the signifi- cance of the vessel state is the directional derivative of the least distinct edge of the vessel segment.

60. The method according to any of the preceding claims 48-59, wherein the characteristics are selected from the significance of the vessel segment, the validity of the starting point, the width of the vessel segment, the direction of the vessel segment, the x- position of the centre point of the vessel, and/or the y- position of the centre point of the vessel, the x- position of a left edge point of the vessel, and/or the y- position of a left edge point of the vessel, and/or the x- position of a right edge point of the vessel, and/or the y- position of a right edge point of the vessel, curvature, colour and/or intensity.

61. The method according to any of the preceding claims 48-60, wherein the characteristics include the width of the vessel segment.

62. The method according to any of the preceding claims 48-61 , wherein at least three characteristics are determined, such as at least four characteristics.

63. The method according to any of the preceding claims 48-62, wherein at least two vessel states are ranked in a priority queue according to their significance.

64. The method according to claim 63, wherein a neighbouring vessel segment is identified in step a3) from a starting point related to the highest ranking vessel state.

65. The method according to any of the preceding claims 48-64, wherein the detected characteristic is weighted with the estimated characteristic at a predetermined constant ratio.

66. The method according to claim 65, wherein the detected width is weighted with the estimated width at a constant ratio, said ratio being in the range of from 0.01 to 1.0, such as in the range of from 0.01 to 0.8, such as in the range of from 0.03 to 0.15, such as in the range of from 0.04 to 0.1.

67. The method according to claim 65, wherein the detected direction is weighted with the estimated direction at a constant ratio, said ratio being in the range of from 0.01 to 1.0, such as in the range of from 0.1 to 0.8, such as in the range of from 0.3 to 0.7, such as in the range of from 0.4 to 0.6.

68. The method according to any of the preceding claims 48-67, wherein the cor- rected starting point is the validated centre of the vessel segment.

69. The method according to any of the preceding claims 48-68, wherein corresponding characteristics for the neighbouring vessel segment is detected in a predetermined look ahead distance from the corrected starting point.

70. The method according to claim 69, wherein the look ahead distance towards a neighbouring segment is a predetermined constant.

71. The method according to claim 69, wherein the look ahead distance towards a neighbouring segment is determined based on the validated characteristics.

72. The method according to claim 71 , wherein the look ahead distance is determined based on the validated width of the vessel segment just determined.

73. The method according to claim 72, wherein the look ahead distance is determined by multiplying a constant with the validated width of the vessel segment just determined.

74. The method according to claim 73, wherein the constant is selected in the range of from 0.01 to 5.0, such as from 0.10 to 2.50, such as from 0.15 to 1.0, such as from 0.2 to 0.8, such as from 0.2 to 0.5.

75. The method according to any of the preceding claims 48-74, wherein the track- ing of a vessel is conducted simultaneously in two opposite directions from the initial starting point of step a2).

76. The method according to any of the preceding claims 47-75, wherein information regarding vessel segments tracked is stored in a memory.

77. The method according to claim 76, wherein the information comprises the significance of the vessel state of each vessel segment tracked.

78. The method according to any of the preceding claims 47-77, wherein a filtered vessel state is determined for each vessel segment.

79. The method according to claim 78, wherein a vessel state is filtered by weighting the vessel state with information regarding previously identified vessel segments in a consecutive sequence.

80. The method according to any of the preceding claims 78-79, comprising, in step b), selecting a consecutive sequence of vessel segments having a vessel state significance and/or a filtered vessel state significance above a predetermined threshold.

81. The method according to any of the preceding claims 77-79, wherein a stop criterion is fulfilled when the filtered significance of the vessel state is below a predetermined exclusion threshold.

82. The method according to any of the preceding claims 47-81 , wherein the consecutive sequence in step b) is identified as a vessel if it comprises at least a predetermined number of vessel segments.

83. The method according to claim 82, wherein the predetermined number is in the range of from 10 to 100, such as in the range of from 25 to 75, such as in the range of from 30 to 50.

84. The method according to claim 82 or 83, wherein the predetermined number is determined as the ratio of the length of the total vessel segments to the average width of the segments.

85. A system for tracking at least one vessel in an image, said image comprising a plurality of vessels, comprising

a) an algorithm for identifying a consecutive sequence of vessel segments until a stop criterion is fulfilled, and an algorithm for each vessel segment determining a vessel state having a vessel state significance, and optionally determing a filtered vessel state,

c) an algorithm for identifying the consecutive sequence of vessel segments selected in step b) as a vessel.