WO2002048949A1

WO2002048949A1 - Method and arrangment for processing digital image information

Info

Publication number: WO2002048949A1
Application number: PCT/SE2001/002702
Authority: WO
Inventors: Björn M. NILSSON
Original assignee: Cellavision Ab
Priority date: 2000-12-15
Filing date: 2001-12-07
Publication date: 2002-06-20
Also published as: AU2002222846A1

Abstract

The invention relates to a method of separating a searched object from a digital microscope image, which comprises a plurality of picture elements and reproduces a biological material. The method is characterized by the steps of selecting in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least one picture element constituting an initial quantity; assigning to at least a first subset of the remaining picture elements in the image object similarity values; calculating object likelihood values, on the basis of said initial quantity and said object similarity values, for at least a subset of the picture elements in the image; and separating the searched object on the basis of the object likelihood values.

Description

METHOD AND ARRANGEMENT FOR PROCESSING DIGITAL IMAGE

INFORMATION

Cross reference to related applications: This application claims benefit from Swedish patent application no SE-0004644-1, filed December 15, 2000, and US provisional patent application no US-60/264350, filed January 26, 2001.

Field of the Invention

The present invention relates to a method of separating a searched object in a digital image according to the preamble of claim 1. The invention also concerns a corresponding arrangement according to the preamble of claim 13 and a digital storage medium comprising a corresponding computer program according to the preamble of claim 14. Background Art A frequent operation in conjunction with the processing of digital images is to separate a searched object in an image. This operation is often referred to as segmenting. Since a digital image, which consists of a number of picture elements (often called pixels) , is an electronic representation of an image, segmenting is usually carried out by means of a computer system. The operation implies that a number of picture elements in the image are selected, which together constitute an electronic representation of an image of the searched object .

As a rule, segmenting precedes other operations, such as image analysis. The purpose of an image analysis can be to classify a segmented object, and therefore it is important for the object to be correctly segmented. In a correctly segmented object essentially the entire object, but essentially nothing of adjoining objects or the background, accompanies the set of picture elements which are selected with the segmenting. An incorrectly executed segmenting can result in a subsequent image analysis producing an incorrect result.

In particular, the present invention is intended to be used in connection with systems for analysis of micro- scope images representing a biological material.

In such a system, for instance a digital microscope image of a blood sample can be used. The digital image then usually shows a number of stained blood cells. The blood cells stand out dark against a brighter background. Typically, an individual blood cell, for instance a white blood cell, should in a segmenting process be separated from the image, i.e. a number of picture elements in the image which jointly represent a blood cell are to be identified. Subsequently, these identified picture ele- ments constitute input data for an image analysis process where the white blood cell is classified. On the basis of classification of a plurality of such blood cells, a di- agnosis of the organism from which the blood sample comes can then be made .

A corresponding, but somewhat more limited, classification can be made of the nucleus of a cell. Then the nucleus is to be separated.

A known method of carrying out segmenting or separating of an object is so-called thresholding. In the above case of blood cells, picture elements in the image, whose intensity is lower than a given threshold value, are said to constitute blood cells. Connected such picture elements are therefore separated by the system that carries out the segmenting.

A system for classification of cells, which uses a thresholding algorithm, is disclosed in US-A-5, 978 , 497. Thresholding algorithms, however, are associated with a few problems. If the object, for instance a first stained white blood cell, which is to be segmented is not free-lying, for instance if a second white blood cell which is also stained is positioned geometrically so close to the first blood cell that they touch each other, an ordinary thresholding algorithm will produce an incorrect result. The result of such a known algorithm will in most cases be that both cells are segmented as one object. A subsequent classification of such a segmented ob- j ect will not produce a usable result.

The problem with adjoining blood cells which are difficult to segment, as mentioned above, is particularly pronounced when analyzing bone marrow specimens since the cell density in such specimens is especially high. However, the problem also occurs in peripheral blood. Summary of the Invention

An object of the present invention is to wholly or partly eliminate the above problems.

This object is achieved by a method of separating a searched object in a digital image according to claim 1, by a corresponding arrangement according to claim 13 and by a digital storage medium comprising a corresponding computer program according to claim 14.

More specifically, the invention then relates, according to a first aspect, to a method of separating a searched object from a digital microscope image which comprises a plurality of picture elements and reproduces a biological material. The method is characterized by the steps of selecting in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least one picture element forming an initial quantity; assigning to at least a first subset of the remaining picture elements in the image object similarity values; calculating object likelihood values based on said initial quantity and said object similarity values, for at least a second subset of the picture elements in the im- age, so that the object likelihood value of a given picture element depends on its object similarity value, its distance to the initial quantity and on the object similarity values of picture elements which are located be- tween the given picture element and the initial quantity; and separating the searched object on the basis of the object likelihood values.

Such a method has been found to function excellently for segmenting objects positioned close to each other. As a rule, there is a thin, in most cases not connected, trace of brighter picture elements between two adjoining darker objects. This trace is formed of a cytoplasmic canal when the objects consist of nuclei, and picture ele- ments corresponding to the trace will to a great extent obtain lower object similarity values than the searched objects. The trace usually adjoins the bright background which is different from the searched objects and therefore can be assumed to be included in the initial quantity. Thanks to the lower object similarity values in the trace and the fact that the calculation starts from a selected quantity outside the searched object, a solution involving comparatively low object likelihood values can be calculated for essentially all picture elements in the trace even if this is interrupted by darker picture elements. This enables more efficient segmenting than in thresholding. In thresholding, an interruption consisting of darker picture elements in the trace would probably result in the two adjoining objects being interpreted as one object.

Preferably, said object likelihood values are calculated by means of a weighted distance transformation. This results in a quick and in most cases sufficiently reliable calculation process

Alternatively, said object likelihood values are calculated by assigning to at least a subset of the pic- ture elements in the initial quantity a first object likelihood value; letting the initial quantity spread in the image by including in the same picture elements preferably adjoining the initial quantity, an included picture element being assigned an object likelihood value which depends on its object similarity value and an object likelihood value of a preferably adjoining, already included picture element . This results in a very accurate segmenting process in particularly difficult circumstances . Preferably, the searched object is a nucleus. The method according to the invention has been found particularly suitable for segmenting nuclei .

According to a preferred embodiment, said at least one first picture element is selected by means of a thresholding operation, so that only bright areas in the image are included in the initially selected quantity. This is a simple and reliable method of selecting a relatively large initial quantity, which results in a quick segmenting process. Preferably, connected areas, of a size below a certain level, in the initially selected quantity are left out from the initial quantity. As a result, a risk of er- rors is avoided, which arises owing to individual bright spots being found in, for example, a nucleus.

The object similarity value of a given picture element can be determined on the basis of its gray scale in- tensity. However, a weighting of the intensity of its primary color components is preferably used. This yields quick and reliable measures of the object similarity of a picture element .

According to an alternative embodiment, the object similarity value of a given picture element is determined on the basis of properties of a set of picture elements in a limited area, in which the given picture element is included. This gives a possibility of determining the object similarity of a picture element on the basis of the texture of an area, i.e. how unevenly stained the picture elements in an area are.

Preferably, a method according to the invention further comprises the steps of identifying picture elements with locally extreme object likelihood values; identify- ing image areas, each of which consists of a number of picture elements which are associated with said locally extreme object likelihood values; processing two such image areas which adjoin each other, as one area if the difference between on the one hand at least one object likelihood value in a picture element in the boundary line between the two image areas and, on the other hand, at least one of the locally extreme object likelihood values associated with the image areas is below a thresh- old value. This gives a possibility of compensating for oversegmentation, i.e. that an object which is to be segmented is in turn divided into several parts .

According to a second aspect, the invention concerns an arrangement for separating a searched object from a digital microscope image which comprises a plurality of picture elements and reproduces a biological material. The arrangement is characterized by means for selecting in the image at least a first picture element, the ap- pearance of which is different from that of expected picture elements in the searched object, said means being arranged in such manner that said at least first picture element forms an initial quantity; means for assigning object similarity values to at least a first subset of the remaining picture elements in the image; means for calculating object likelihood values, on the basis of said initial quantity and said object similarity values, for at least a second subset of the picture elements in the image, so that the object likelihood value of a given picture element depends on its object similarity value, its distance to the initial quantity and on the object similarity values of picture elements which are located between the given picture element and the initial quantity; and means for separating the searched object on the basis of the object likelihood values.

The arrangement gives advantages corresponding to those of the method and can also be varied similarly. According to a third aspect, the invention relates to a digital storage medium comprising a computer program for separating a searched object from a digital microscope image which comprises a plurality of picture ele- ments and reproduces a biological material. The storage medium is characterized by instructions corresponding to the steps of selecting in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least one picture element forming an initial quantity; assigning to at least a first subset of the remaining picture elements in the image object similarity values; calculating object likelihood values, on the basis of said initial quantity and said object similarity values, for at least a subset of the picture elements in the image, so that the object likelihood value of a given picture element depends on its object similarity value, its distance to the initial quantity and on the object similarity values of picture elements which are located between the given picture element and the initial quantity; and separating the searched object on the basis of the object likelihood values.

The computer program gives advantages corresponding to those of the method and can also be varied similarly. Brief Description of the Figures

Figs la and lb show a digital image of four adjoining, stained white blood cells surrounded by a number of red blood cells. Fig. 2 shows an initial quantity of picture elements selected in Fig. la according to the invention.

Fig. 3 indicates, according to an embodiment of the invention, determined object similarity values for pic- ture elements in the digital image in Fig. la.

Fig. 4 indicates, according to an embodiment of the invention, determined object likelihood values for picture elements in the digital image in Fig. la.

Fig. 5 shows the digital image in Fig. la, where in- teresting areas have been identified on the basis of the object likelihood values shown in Fig. 4.

Fig. 6 shows the interesting areas in Fig. 5 after combining them according to the invention.

Figs 7a and 7b show an arrangement according to the invention.

Fig. 8 is a flow chart of a method according to the invention. Description of Preferred Embodiments

Fig. la shows a digital image of four adjoining, stained white bloods cells surrounded by a number of red blood cells. Fig. lb shows silhouettes of elements included in Fig. la, with reference numerals. Figs la and lb are to be seen together.

In the center of the digital image in Fig. la there are four white blood cells (often called white corpuscles) 1-4. They are stained with a dye in order to appear more clearly in the digital image. Also a set of red blood cells (designated R) are to be seen in the image. A large part of the cross-sectional surfaces of the white blood cells 1-4 is occupied by their respective nuclei (whereas red blood cells normally have no nuclei) . However, it is possible to see, in some of the cells, a boundary line between the nuclei and their surrounding cytoplasm (the boundary lines are indicated by dashed lines in Fig. lb) .

A thresholding segmentation of prior art type, that tried to segment nuclei in white blood cells, would prob- ably produce an incorrect result if it were applied to the image in Fig. la. This is due to the fact that the white blood cells in the image are positioned so close to each other that they have no background between them and that also their nuclei adjoin each other. By a method according to the invention, one tries to separate an object which preferably is a nucleus in a white blood cell. However, also an entire white blood cell, a red blood cell or another object in a biological, preferably microscoped, material can be involved. Exam- pies of biological materials can be blood, bone marrow, cervical smears, or tissues, such as liver tissue.

In connection with the images in Figs 2-6, it will now be described how a segmenting process according to the invention is used to separate or segment the white blood cells in Fig. la.

A method according to the invention can be generally described as follows . An initial quantity of picture elements in an image is selected. These are selected in such manner that with some degree of certainty they do not reproduce the searched object. They can be assigned object likelihood values reflecting this.

Object similarity values which are assigned at least a first subset of picture elements in the image are also calculated. A given picture element is then assigned an object similarity value which indicates how similar the given picture element (or a group of picture elements in a small area, in which area the given picture element is included) is to expected picture elements in the searched object.

Subsequently object likelihood values of at least a second subset of picture elements in the image are calculated. This is carried out on the basis of the initial quantity and the determined object similarity values.

The assigned object likelihood value of a given picture element expresses how probable it is that a given picture element, in its context, reproduces a searched object. The object likelihood value preferably depends on the object similarity value of the given picture element, on how close to the initially selected quantity it is located and on object similarity values of picture elements positioned between the given picture element and the initially selected quantity. On the basis of these object likelihood values, the searched object can then be separated. This can be carried out in such manner that picture elements with object likelihood values above a certain level are assumed to be included in objects of the searched type and are processed accordingly. SELECTION OF INITIAL QUANTITY

First an initial quantity is selected by selecting at least one picture element in the image, the appearance of which is different from that of the expected picture element in the searched object.

Fig. 2 shows an initial quantity of picture elements selected in Fig. la according to the invention. The pic- ture elements selected in Fig. la are in the image in Fig. 2 shown as white whereas the non-selected picture elements are shown as black.

The selected picture elements are selected according to a criterion so that they are visually different from expected picture elements in the searched object. In the case shown, nuclei in stained white blood cells are to be separated. These are significantly darker than the background. It is therefore convenient in this case to set up a criterion so that bright picture elements are included in the initially selected quantity. It is advantageous to place the demands of the criterion high so that with a high degree of certainty no picture elements inside a searched object will be included in the selected quantity. Otherwise, the method may produce an incorrect result. In the shown example, picture elements can be selected whose intensity is lower than 0.5 on a scale where black is 1 and white is 0. In the shown example, it may also be advantageous to exclude picture elements belonging to connected bright areas which are smaller than a certain size since such areas may in some cases be found in a nucleus. The ini- tially selected quantity then includes merely large connected chunks of bright picture elements.

The criterion used can also be based on a certain color content or on the texture of a group of picture elements . DETERMINATION OF OBJECT SIMILARITY VALUES

When the initial quantity has been selected, object similarity values are determined, for at least a first subset of the remaining picture elements. The first subset can, but need not, comprise all picture elements out- side the initially selected quantity. It can in principle also comprise picture elements in the selected quantity. The object similarity value of a picture element is a measure of how similar the picture element (or a group of picture elements in which the picture element is in- eluded) is to expected picture elements in the searched object. A bright picture element is probably not included in a searched dark object and is therefore given a low object similarity value, whereas a dark picture element obtains a high object similarity value. The object similarity value of a given picture element can be determined as a function of its gray scale intensity. This function can, but need not, be linear. The object similarity value of a given picture element can also be determined as a weighting of the intensity of its primary color components, also referred to as RGB projection. Nor does this have to be linear. The object similarity value of a given picture element can also be determined on the basis of properties of a set of picture elements in a limited area, in which the given picture element is included. The given picture element can, but need not, be included in this set. It is then also possible to assign an object similarity value on the basis of the texture of the area.

Fig. 3 indicates, according to an embodiment of the invention, determined object similarity values for picture elements in the digital image in Fig. la. The fact that a picture element in Fig. 3 is dark means that the corresponding picture element in Fig. la has a high object similarity value. The fact that a picture element in Fig. 3 is bright means that the corresponding picture element in Fig. la has a low object similarity value.

CALCULATION OF OBJECT LIKELIHOOD VALUES

Subsequently object likelihood values are determined for at least a second subset of picture elements in the image on the basis of the initial quantity and the determined object similarity values. The object likelihood value of a given picture element is a measure of how probable it is that the picture element reproduces an ob- ject of the searched type. An object likelihood value can be a real number of an arbitrary magnitude.

A method of separating, on the basis of the initially selected quantity and the determined object similarity values, a searched object, such as a blood cell, from the image is using a model of contour spreading, preferably a so-called fast marching model. The method can be carried out by letting the initially selected quantity of picture elements spread by including in the selected quantity picture elements preferably adjoining the selected quantity. Preferably, at least some of the picture elements in the initial quantity are assigned a first, basic object likelihood value. A picture element included in the selected quantity is then assigned an object likelihood value which depends on and preferably is a function, for instance the sum, of the picture element's own object similarity value and the object likelihood value of an adjoining, already included picture element. If the function is a sum, the later included picture element will obtain higher and higher object likelihood values. Suitably, the one of the picture elements, adjoining the selected quantity, which will result in the lowest object likelihood value, will be included first in the selected quantity.

The above method can be resembled to a fire spreading over a forest area. The areas which are just catching fire then form contours round already burning areas . These contours will move more quickly over dry (and thus inflammable) areas and more slowly over wetter areas.

Transferred to the example of white blood cell nuclei, this model of contour spreading behaves in such manner that the "fire" starting in areas which with a high degree of certainty do not reproduce a nucleus, spreads very quickly over bright (dry = low object similarity values) areas (which gives low object likelihood values) and only slowly over dark (moist = high object similarity values) areas, which gives quickly increasing object likelihood values. The object likelihood values can then be said to be the time from the image first "catching fire". The object likelihood value of a given picture element depends on its object similarity value, its distance to the initial quantity and on object similarity values of picture elements which are positioned between the given picture element and the initial quantity. It can be intuitively understood that the fire will "sneak" into the brighter trace between the nuclei . Picture elements between the given picture element and the initial quantity may also refer to picture elements slightly outside the shortest path between the given picture element and a picture element in the initial quantity. This is equivalent to the "fire" following a highly inflammable, but not necessarily straight, trace between the nuclei .

The result of such a method according to an embodiment of the invention is evident from Fig. 4. This Figure indicates determined object likelihood values of the picture elements in Fig. la. The brighter a picture element is in Fig. 4, the higher its object likelihood value. It is apparent that the object likelihood values are lower in the area where the nuclei adjoin each other than in the center of a nucleus.

It is also possible to approximate the object likelihood values of the picture elements by using a so- called distance transformation. Such a transformation is described in detail in "Distance Transformations in Digi tal Images" by Gunilla Borgefors, Computer Vision Graphics and Image Processing 34 . 344 -371 (1986) . Such a transformation can be modified in such manner that it also takes object similarity values, not only euclidian distances, into consideration.

Using a distance transformation it is possible to calculate, for each picture element in an image, the distance to a given zero level. This is carried out by repeatedly applying so-called masks to the image, as will be described below. In connection with the present invention, the masks, when they are applied to a picture element in the image, should besides be weighted with regard to the object similarity value of this picture element. A simple example of a matrix is shown below, which has its equivalence in a digital image with 4x4 picture elements. In the image, a number of picture elements have been selected to form an initial quantity. Corresponding elements in the matrix below have been assigned the ob- ject likelihood value zero (0) . The example is shown with integers, but also floating-point numbers can be used. The remaining elements in the matrix are assigned the value MAXINT (M) , which in these numerical calculations roughly corresponds to infinity:

0 0 0 0

0 M M M

0 M 0 0

For a satisfactory result when, as will be shown below, a 3x3 mask is applied to the image, a frame of MAXINT values is placed round the matrix and forms a modified matrix:

M M M M M M

M 0 0 0 0 M

M 0 M M M M

M 0 M 0 0 M

M M M M M M Subsequently a forwardly directed mask is applied to the modified matrix. This is applied first in the 2,2 element in the modified matrix, (1,1 in the original), then in 2,3; 2,4; 2,5; so that it is applied to all elements in the first row of the original matrix. Subsequently the mask runs over the second, third and fourth rows of the original matrix in the same way, i. e. from the left to the right.

The forwardly directed mask is as follows: 4 3 4

3 0 M

M M M

The selection of the values 3 and 4 in the mask above depends mainly on the fact that 4/3 is a good approximation of "V2, which is the length of hypotenuse in a right- angled triangle with the side 1. When calculating using floating-point numbers, masks with better approximations of "V2 can be used. When the mask is applied over a given element in the matrix, it is placed over the matrix in such manner that the central element of the mask is placed over the given element of the matrix. The elements of the mask are then summed with the respective elements of the underlying matrix. For the element 2,2 in the modified matrix, the sums will be as follows:

4 + M 3 +M 4 + M

3 + M 0+0 M + 0

M + M M + 0 M + M The matrix element to which the mask is applied is then assigned the value of the smallest of these sums, i.e. 0+0=0. In fact, the element 2,2 in the modified matrix already had this value. The now assigned value is a measure of the distance of the matrix element to the zero level and this distance is in fact zero. In the same way, the values of the elements 2,3; 2,4 and 2,5 are then de- termined to be zero. In computer processing, summations where M is included are normally not carried out.

When the mask is eventually applied to the 3,3 element in the modified matrix, the sums will be as follows:

4+0 3+0 4+0 3 + 0 0+M M +M

M+ 0 M+M M +M

The smallest of these sums is 3+0=3, and therefore the 3,3 element in the modified matrix is assigned the value 3, which is a measure of the distance between this element and the zero level. By letting the mask run over the modified matrix, more and more elements in the matrix are assigned values which are different from MAXINT and constitute measures of the respective distances between the elements and the zero level.

When the forwardly directed mask has been applied to the modified matrix as above, also a rearwardly directed mask is applied to the matrix. This should run over the matrix from the right to the left, row by row, from below upwards. The rearwardly directed matrix is as follows:

M M M

M 0 3

4 3 4

The result which is obtained by the above example, however, merely gives information about the approximated euclidian distance between the initially selected quantity and a matrix element (and, thus, a corresponding picture element) . According to the invention, the mask (forwardly or rearwardly directed) should therefore, when applied to a picture element, be weighted as regards the object similarity value of the picture element. This can be carried out by the elements of the mask being multiplied by the object similarity value li .j of the picture element i,j. In the Example with the matrix element 3,3 above the sums will be as follows:

/₃₎₃ - 4 + 0 /_3ι3 - 3 + 0 /_3>3 - 4 + 0 /_3>3 - 3 + 0 /_3>3 - 0 + M /_3i3 - M + M l_{3 3} - M + 0 l ₃ - M + M l_{3 3} - M + M The value which is assigned to a matrix element will now in the first place not be a measure of the distance to the initially selected quantity, but a measure of the object similarity value of the corresponding picture element and of object similarity values of picture elements which are positioned between the picture element in question and the initially selected quantity.

The values of the matrix elements calculated as described above are suitable object likelihood values of their corresponding picture elements. USE OF OBJECT LIKELIHOOD VALUES

When the object likelihood values have been determined, it is possible to identify locally extremes (usually maxima), i.e. picture elements having higher object likelihood values than all their neighbors. For each such local extreme, an associated area can be identified. This is conveniently carried out in such manner that picture elements in the vicinity of a local extreme, which have lower and lower object likelihood values the further away from the local extreme they are positioned, are associated with the local extreme. This can be combined with a threshold level regarding object similarity value or object likelihood value.

Fig. 5 shows the digital image in Fig. la, where such areas associated with local extremes have been identified. As is evident, a certain, so-called hypersegmen- tation has taken place, i.e. the nuclei (which should in fact be segmented) have each in turn been divided into several parts. These areas are therefore combined if certain conditions are satisfied. Two adjoining areas are combined if the difference between, on the one hand, an object likelihood value at the boundary line between the areas and, on the other hand, one of the extreme object likelihood values associated with the areas is below a threshold value .

Fig. 6 shows the areas in Fig. 5 after such combina- tion of areas. The areas shown in Fig. 6 constitute, according to circumstances in the image, excellent segmentations of the nuclei in Fig. la. The areas shown in Fig. 6 can be input data for an image analyzing process, such as classification. Fig. 7a illustrates a system with an arrangement according to the invention. The system comprises a digital microscope 71, which is connected to a computer system 72 The digital microscope 71 supplies digital microscope im- ages in black-and-white or color to the computer system 72. Before the digital images are supplied to the computer system 72, some kind of digital processing can be performed. The computer system 72 can be integrated in the microscope 71.

The computer system 72 comprises means 73-77 for carrying out the steps included in the segmentation process described above.

A means 73 for selecting, based on the digital image, an initial quantity is included. The output signal of this means is an indication of which picture elements have been selected in the image, cf . Fig. 2.

Moreover, a means 74 for determining object similarity values of the picture elements in the digital image is included, as mentioned above. The output signal of this means is an indication of the object similarity values of picture elements included in the digital image, cf. Fig. 3.

The respective output signals of these two means are inputted into a means for determining object likelihood values 75. This means determines, on the basis of its input signals, object likelihood values of picture elements in the image, which object likelihood values are out- putted, cf. Fig. 4. On the basis of the object likelihood values, a means for separating 76 carries out segmentation of searched objects in the image. The areas shown in Fig. 5 constitute an intermediate product . The segmented ob- jects, cf . Fig. 6, can then be analyzed in a classification unit 77.

The above-mentioned means 73-77 can generally consist of the computer system, provided with a computer program for carrying steps according to the method. This computer program can be stored separately on a digital storage medium. It is, however, also possible to accomplish at least some of the above means in hardware, such as ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array) circuits.

It is also possible to carry out calculation steps in a distributed manner. The segmentation process need not be carried out in situ adjacent to the microscope. The entire method, or parts thereof, can be carried out, for example, in an Internet server, positioned at a distance from the microscope.

Fig. 8 is a flow chart of a method 80, according to the invention, of separating a searched object in a digital image . In a first step 81, at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, is then selected in the image. This at least one picture element constitutes an initial quantity. In a second step 82, at least a first subset of the remaining picture elements in the image is assigned object similarity values. In a third step 83, object likelihood values are calculated for at least a subset of the picture elements in the image, on the basis of the initial quantity and the object similarity values. In a fourth step 84, the searched object is separated on the basis of the object likelihood values.

The invention is not limited to the embodiments described above but can be varied within the scope of the appended claims. In summary, the invention relates to a method of separating a searched object from a digital microscope image, which comprises a plurality of picture elements and reproduces a biological material . The method is characterized by the steps of selecting in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least one picture element constituting an initial quantity; assigning to at least a first subset of the remaining picture elements in the image object similarity values; calculating object like- lihood values, for at least a subset of the picture elements in the image, on the basis of said initial quantity and said object similarity values, so that the object likelihood value of a given picture element is dependent on its object similarity value, its distance to the ini- tial quantity and on object similarity values of picture elements which are positioned between the given picture element and the initial quantity; and separating the searched object on the basis of the object likelihood values .

Claims

CLAIMS 1. A method of separating a searched object from a digital microscope image which comprises a plurality of picture elements and reproduces a biological material, c h a r a c t e r i z e d by the steps of selecting (81) in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least first picture element forming an initial quantity; assigning (82) to at least a first subset of the remaining picture elements in the image object similarity values; calculating (83) object likelihood values, based on said initial quantity and said object similarity values, for at least a second subset of the picture elements in the image, so that the object likelihood value of a given picture element is dependent on its object similarity value, its distance to the initial quantity and on object similarity values of picture elements which are positioned between the given picture element and the initial quantity; and separating (84) the searched object on the basis of the object likelihood values.

2. A method as claimed in claim 1, wherein said object likelihood values are calculated by means of a weighted distance transformation.

3. A method as claimed in claim 1, wherein said object likelihood values are calculated by assigning to at least a subset of the picture elements in the initial quantity a first object likelihood value; letting the initial quantity spread in the image by including picture elements in the initial quantity, an included picture element being assigned an object likelihood value which depends on its object similarity value and an object likelihood value of an already included picture element.

4. A method as claimed in any of claims 1-3, wherein said searched object is a nucleus.

5. A method as claimed in any of claims 1-3, wherein said searched object is a white blood cell.

6. A method as claimed in any of claims 1-3, wherein said searched object is a red blood cell.

7. A method as claimed in any one of the preceding claims, wherein said at least first picture element is selected using a thresholding operation so that only bright areas in the image are included in the initially selected quantity.

8. A method as claimed in claim 7, wherein connected areas in the initially selected quantity of a size below a certain level are excluded from the initial quantity.

9. A method as claimed in any one of the preceding claims, wherein the object similarity value of a given picture element is determined on the basis of its gray scale intensity.

10. A method as claimed in any one of claims 1-8, wherein the object similarity value of a given picture element is determined on the basis of weighting of the intensity of its primary color components.

11. A method as claimed in any one of claims 1-8, wherein the object similarity value of a given picture element is determined on the basis of properties of a set of picture element in a limited area, in which the given picture element is included.

12. A method as claimed in any one of the preceding claims, further comprising the steps of identifying picture elements with locally extreme object likelihood values; identifying image areas, each of which consists of a number of picture elements which are associated with said locally extreme object likelihood values; processing two such image areas, which adjoin each other, as one area if the difference between on the one hand at least one object likelihood value of a picture element in the boundary line between the two image areas and, on the other hand, at least one of the locally extreme object likelihood values associated with the image areas is below a threshold value.

13. An arrangement for separating a searched object from a digital microscope image which comprises a plurality of picture elements and reproduces a biological material, chara c t e r i z e d by means for selecting (73) in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said means being arranged in such manner that said at least one picture element forms an initial quantity; means for assigning (74) object similarity values to at least a first subset of the remaining picture elements in the image; - means for calculating (75) object likelihood values, on the basis of said initial quantity and said object similarity values, for at least a subset of the picture elements in the image, so that the object likelihood value of a given picture element is dependent on its ob- ject similarity value, its distance to the initial quantity and on object similarity values of picture elements which are positioned between the given picture element and the initial quantity; and means for separating (76) the searched object on the basis of the object likelihood values.

14. A digital storage medium comprising a computer program for separating a searched object from a digital microscope image which comprises a plurality of picture elements and reproduces a biological material, c h a r - a c t e r i z e d by instructions corresponding to the steps of selecting in the image at least a first picture element, the appearance of which is different from that of expected picture elements in the searched object, said at least one picture element forming an initial quantity; assigning to at least a first subset of the remaining picture elements in the image object similarity values; calculating object likelihood values, on the basis of said initial quantity and said object similarity values, for at least a subset of the picture elements in the image, so that the object likelihood value of a given picture element is dependent on its object similarity value, its distance to the initial quantity and on object similarity values of picture elements which are positioned between the given picture element and the initial quantity; and - separating the searched object on the basis of the object likelihood values.