DE3917760C1 - Microscope with video scanning camera - for automatic examination of preparations e.g. slides with an object recognition system - Google Patents

Abstract

An object recognition system that is used for examination of preparations and is carried out in real time. The system uses digital electronic image signals obtd. from a scanning camera coupled to a microscope. The system allows specific areas to bee examined based upon edge detection. The electronic system has a large number of processing elements (E) arranged in groups (G) within a defined framework (R). The outputs of the groups connect with an arithmetic adder unit (A) to provide a value for comparison with a defined window selection )value. ADVANTAGE - Allows automatic examinationof specific areas

Description

The invention relates to a method for recognizing objects, the properties of which are present in the form of digitized electronic image signals, in video real time, in particular for the automatic screening of cytological preparations, an evaluation frame (R) relating to a section of the digital image signal Field, the evaluation frame has evaluation elements (E) arranged in groups and the groups (G) essentially define radially emanating rays originating from a coordinate origin.

The screening routine in cytodiagnostics should be automated to such an extent that the trained personnel only have to look at the interesting areas of a preparation; the preparation is scanned using a video camera and a microscope with an XY table. If an atypical cell is recognized, the machine saves its coordinates so that the specialist can later make a diagnosis for the image situation in question. Sufficiently high processing speeds can only be achieved by machines that examine only the most striking of several criteria for cell atypia: an enlarged cell nucleus. DE-OS 37 08 795 describes a method which uses a comparison mask to select objects according to shape classes, a rather high degree of misinterpretation ( FIG. 3, mfp and mfn) being unavoidable due to the naturally large fluctuation ranges of the cell nucleus geometries .

The method of restricting oneself to one criterion of the cell nucleus surface already means restricting it to a certain type of cell atypia, which is justified if this criterion is also recognized with a high degree of certainty. Atypically enlarged cell nuclei generally deviate more or less from the round shape. They are statistically oriented on the slide; masking must be done with a circle-like mask, so that the hit rate averaged over all cores is maximum. A method with a high degree of false negative results (= poor sensitivity; Fig. 3, mfn) is hardly justifiable, especially for medical diagnostic purposes . Individual deformed atypical cell nuclei cannot show full agreement with the round mask and a lowering of the comparison threshold is necessary for their detection. A compulsion to lower the threshold even more results from the influence of the vicinity of a cell nucleus on the match result. If the sensitivity is increased by lowering the threshold value, a large number of objects that are not atypical cells are recognized. These false positive alarms (= poor selectivity; Fig. 3, mfp) endanger the original goal of the automatic screening, to relieve the staff.

The invention is intended to solve the problem of measuring the area of the digital electronic image signals of objects (sig) largely independently of the special shape of the object boundary; in addition, the area dimension (fmz) determined by the invention should be largely independent of the immediate surroundings of the object to be measured. The invention is intended to carry out this area measurement so quickly and in such a data format that it enables automatic selection of objects, the incorrect rates, i.e. the non-recognition of sought objects ( Fig. 2, rfn) and the erroneous recognition of objects not sought ( Fig. 2, rfp) , should be lower than in the above-mentioned method ( FIG. 3, mfn and mfp) .

For this purpose, according to the characterizing features of claim 1 Assessment framework used that has no restrictive assumptions yet anticipates properties of the object. The in the assessment framework summarized evaluation elements are sensitive to detection signals of the image section currently being viewed.

A detection signal can have exactly two states: 0 or 1 (or also low-level or high-level). One of those two potential Values is considered a positive detection signal; which one both depend on the circuit logic, on the other on the contrast between the object and the background.

Assessment elements of each group only contribute to the measurement result at (= are not desensitized) if all are radially further inside lying evaluation elements of the same group a positive detection signal exhibit.

With a certain circuit logic, for example 0 for one bright pixel and 1 for a dark one is the definition of positive only depends on the contrast between the object and the background: If the object z. B. by lower brightness of his background stands out, is the example given above Circuit logic of state 1 defined as positive.  

In the simplest case of a 1-bit brightness-resolved image signal, this can be done positive detection signal z. B. be a digital high-level, which is then for "darker Surface area ", then the negative detection signal would be corresponding low-level for "light background". The entirety of the detection signals forms a pixel pattern at any point in time, which is a digitized analogue of the Picture section shows the picture situation to be analyzed. Several evaluation elements are combined into a group, within which logical links the Evaluation elements are made. As a result, the answer of a single evaluation element depends on a detection signal depending on which detection signals the others Received rating items from the same group. The special arrangement of the Groups occur radially in a way that is unusual for digital electronic circuits, which ensures that the logical links are effective across the board.

The characteristic features of the invention result in a radial coordinate system implemented in hardware, the direction vectors of which diverging at certain angles are formed by the groups (G) . The invention detects the boundary curve of an object by determining its location vectors. The sensitive subgroups serve as location vectors. The task can be solved by knowing the edge of the object; disruptive positive detection signals (sig) from the neighborhood of the object can be suppressed because it is known where this neighborhood begins. The area size follows by summing the detection signals (sig) within the edge using the output values of the evaluation elements (E) .

Claim 2 specifies an advantageous embodiment for the formation of the subgroups. The hierarchical hierarchy provides a directional link between the Assessment elements; exactly then is z. B. such an evaluation element is sensitive, if it is not preceded by an evaluation element of the same group that is a negative one Detection signal received.

Claim 3 gives an advantageous embodiment for minimizing the area measurement error: since the beams defined by the groups diverge radially, the radiance increases towards the coordinate origin. In terms of circuitry, it would be very complex, the evaluation elements accordingly different sizes Assign patches; it is better to leave overlays of Be, for example evaluation elements and weight their response to a positive detection signal according to the number of evaluations receiving the same signal together elements either proportionately or by one of the evaluation elements gives full answer and the rest are silent, but possibly hierarchical do not desensitize subordinate assessment elements.

Claim 4 specifies an advantageous further development to avoid misinterpretations: only objects lying completely within the evaluation framework can be measured reliably. To determine whether this condition is met, for. B. can be checked whether all groups contain at least one desensitized element (E) . If not, a marker is output; a marking line (mrk) is queried for the marking, and if the answer is positive, e.g. B. the measurement result in question can be zeroed.

Claim 5 specifies an advantageous embodiment for realizations of the method, who use physically or logically separate electronic components for the groups. Then the area measure is determined by adding the group results, whereby the latter already contain the weighting according to claim 3.  

Claim 6 specifies an advantageous embodiment. PAL modules offer one of the large number of address inputs, which achieves high area resolution can be. In addition, PAL modules are programmed with logic functions, whereby the integration density is used in the use according to the invention becomes. After all, PAL modules are not programmed volatile. The above the mask filter method mentioned requires transiently programmed comparison modules and multiplexers if it does not want to lose the variability of the mask. The invention is, on the other hand, intrinsically variable in shape due to the absence of masking.

Claim 7 specifies an advantageous further development, which adapts the measuring method to the needs of existing techniques, the advantages of the invention being largely retained; in particular, filtering with windows means greater flexibility than filtering with a single threshold. It is possible to set several windows (F) and to arrange the examined objects in several classes during a single work step, as z. B. is required in production monitoring for sugar production. However, by setting windows, the original information of the area dimension is reduced, which can be accepted if you have to carry out a large number of measurements in short times and manage them with little electronics. There is another advantage of the invention over the mask filter method when it is used for single-stage size selection; Instead of having to appropriately program mask shape, area and threshold value, it is sufficient to specify the window limits in the method according to the invention. This also means less susceptibility to incorrect operation, because in order to find the right mask / threshold combination for a size selection problem, not only a lot of skill and experience is usually required, but also careful testing of the possible options, since the interactions of the mask, threshold, scatter of the object shape, scatter of the object size and disturbing object surroundings cannot be immediately overlooked. In contrast, specifying an upper and a lower area boundary is much easier.

The method according to the invention is preferably carried out by a device realized according to claim 8. A logical circuit does the desensitization, either through its wiring (Hardware) or through their programming (software) or through a combination of both.

Such a combination is used in the device according to claim 9 through the interconnection and programming of PAL components turns.  

A particular advantage of the invention can be seen in the fact that essentially the same standard methods for pre- and post-processing can be used as in the mask filtering mentioned above. Analog video signals (vid) are e.g. B. discriminated threshold and converted with suitable series-parallel converters (SP) into a pixel field from detection signals (sig) , which is now wired according to the type described in claim 1 with the evaluation elements (E) . This also ensures the synchronism between the video write beam and the evaluation frame (R) . The area dimension (fmz) appearing at the output of the evaluation frame (R) or, in the case of discretely realized groups (G), at the output of the adder (A ) can be stored or processed further by calculation or can be fed to a window comparator (K) .

The following embodiment explains the invention with reference to the drawing ( Fig. 1):

Since the special wiring of the components is crucial in the invention, a wiring diagram ( FIG. 1) is shown. The 1-bit brightness-resolved video signal (vid) is converted into an 8 × 8 bit pixel field by means of a series-parallel converter (SP) ; 64 detection signals (sig) are available. Eight of them are grouped as bold line buses. These detection signals (sig) are supplied to the evaluation elements (E) in accordance with the invention, for which purpose they are successively removed from the line buses at the respectively suitable locations. In the example there are 112 evaluation elements (E) , which form the evaluation framework (R) in 28 groups (G) with 4 evaluation elements (E) each. The difference between the number of detection signals (sig) and the evaluation elements (E) results from the area normalization of the detection signals (sig) by the uniform pixel clock. Because of the radial arrangement, the assigned area for each evaluation element (E) grows towards the outer limit of the evaluation frame (R) . To compensate for this, interconnections of several evaluation elements (E) from different groups (G) are implemented further inside; these overlaps are taken into account by programming the groups (G) accordingly. The further processing electronics are shown as a block diagram within the inner limit of the evaluation frame (R) : the output signals (gmz) of the groups (G) are added to the area dimension (fmz) in an adder (A) . This is compared in the window comparator (K) with the two window barriers that are preset on the window selector (F) . 6 groups were selected as groups (G) according to claim.

For the sake of clarity, it is assumed here that each PAL module contains exactly one group, that is, it has 4 address inputs and at least 3 outputs. The wiring diagram changes only slightly if two radially opposite groups (G) are combined in a PAL with 8 address inputs and at least 4 outputs: half of the groups (G) have no group outputs to be wired to the adder (A) , because the PAL as a whole provides a 4-bit output value in which the pixel pattern of both groups (G) implemented in it is evaluated and this output value is wired to the adder (A) in a 4-bit line bus. In the example, however, a 3-bit wide bus from each group (G) leads to the adder (A) . The programming of the PAL modules without weighting is done as follows (Tab. 1): where x can be any recognition signal, which achieves the desensitization of the address inputs, which are denoted by x . Positive detection signals are set bits, ie 1, negative detection signals are 0. In order to include the weighting, only the output values have to be selected individually for each PAL, the address pattern of the table given above remaining the same. For this only an example is given, in which a PAL has its bit 0 conductively connected to bit 0 of the PAL shown in the table above (Tab. 1). Then bit 0 of this PAL component to be weighted must not make any contribution to the area dimension, but the desensitization property must be fully preserved. This PAL would then have to be programmed as follows (Tab. 2):

The corresponding considerations apply to all PAL modules with their respective conductors associated address inputs.  

The special marking for oversized objects is that the highest bit of the relevant group output signal is set when all evaluation elements (E) of a group (G) , ie all addresses of a PAL module, receive a positive detection signal (sig) ; then the OR operation of the highest bits of all groups (G) is formed in the adder (A) . The graphic representation of this part is dispensed with; the formation of the special marking is simply regarded as a second function of the adding unit (A) . This special marking is placed on the marking line (mrk) . In the example, the marker line (mrk) is AND-linked (&) with the detection signal (det) , which is set by the window comparator (K) if the area dimension (fmz) is within the limits as preset on the window selector (F) . The result of this link is given to the alarm line (alr) . An object triggers a positive alarm signal that can be used as a trigger to save its coordinates if the determined area dimension (fmz) lies within the window area that was preset with the window selector (F) and if none of the 28 PAL Blocks receive a positive detection signal (sig) at all 4 address inputs.

The mode of operation and properties of the invention were studied using a computer simulation program and compared under the same conditions with the likewise simulated mask method mentioned above. Based on the experience gained from working with a microscope, a library of over 6000 image situations was created and scanned using both methods. In FIGS. 2 and 3 the results are shown as a degree of false rates depending on the set threshold fidelity. Both methods were tested with a resolution of 16 × 16 pixels. Fig. 3 relates to the mask ( 200 ) shown in DE-OS 37 08 795, but this is optimized and rounded off in the context of the higher resolution. A threshold fidelity of 100% means that the threshold value ( FIG. 3) requires that all pixels match the 256 mask pixels for a detection. Fig. 2 shows the entsprendenden results from the inventive process; here 100% threshold accuracy means that the window boundaries ( FIG. 2) have been placed exactly on the edges of the definition range for atypical cell sizes; 72 square micrometers was chosen as the lower edge of this definition range and 250 square micrometers as the upper edge. The nuclei of healthy cells are approximately 30 square micrometers in size.

Lower threshold fidelity means lowering the detection threshold ( FIG. 3) or enlarging the window ( FIG. 2), which in both cases leads to an enlargement of the sensitive area. A false rate of 1 means that none of the atypical cells (mfn and rfn) or that all objects that are not atypical cells (mfp and rfp) triggered a detection alarm . False rate curves (mfp, mfn, rfp, rfn) are aimed for as close as possible to zero. A threshold setting, in which the rate of false negative results is minimal, can be considered as the operating point of a method. From this follows the level of false positive alarms to be accepted at the working point (rap, map) . The mask process does not find all atypical cells from the library if the threshold is not met. The working point is below 80% threshold compliance (map) , which leads to a very high degree of false positive results (mfp) above 0.4. The invention enables an operating point at 99% (rap) threshold fidelity with a low degree of false positive results (rfp) below 0.1.

Claims (9)

1. Method for the detection of objects, the properties of which are present in the form of digitized electronic image signals, in video real time, in particular for the automatic screening of cytological preparations, whereby
an evaluation frame (R) is placed over a section of the digital image signal field,
the evaluation framework has evaluation elements (E) arranged in groups and
the groups (G) essentially define rays emanating radially from a coordinate origin,
characterized in that
Subgroups of evaluation elements (E) in the direction of these beams are desensitized depending on the digital pixel values (sig) of the image situation and non-desensitized evaluation elements provide initial values from which the shape and / or size of the surface of the object are determined. 2. The method according to claim 1, characterized in that the desensitization of the subgroups is carried out as a function of a hierarchical ranking of the evaluation elements (E) within the respective group (G) . 3. The method according to claim 1 or 2, characterized in that the output values of the evaluation elements (E) are weighted individually. 4. The method according to any one of claims 1 to 3, characterized in that the result of an area measurement is particularly marked (mrk) when the object in question protrudes beyond the limits of the evaluation frame (R) . 5. The method according to any one of claims 1 to 4, characterized in that the output values or the weighted output values of the evaluation elements (E) each group (G) combined to form a group value and the group values added to the area measure (fmz) of the entire object (A ) . 6. The method according to any one of claims 1 to 5, characterized in that
that the groups (G) are realized in PAL modules, one PAL forming one or more groups (G) ,
that the evaluation elements (E) are implemented as address inputs of the PAL modules
and that the desensitization and, if necessary, the weighting and special marking (mrk) can be achieved by appropriate programming of the PAL modules. 7. The method according to any one of claims 1 to 6, characterized in that
that with 2 n predeterminable thresholds (F) n window areas for the area dimension (fmz) are defined and
that all measured values of the area measurement (fmz) are then checked (K) whether they are within one of the window areas and, if so, in which they are. 8. Device for performing the method according to one of claims 1 to 7, characterized in that
the outputs of a series-parallel converter are connected to inputs of a logic circuit,
the inputs of the logic circuit realize the evaluation elements of the evaluation framework, and
the wiring and / or the programming of the logic circuit is designed such that groups are formed and desensitization is carried out within these groups. 9. The device according to claim 8, characterized in that the logic circuit consists of one or more PAL components, in which the address inputs of the PAL module or modules the evaluation elements of the evaluation framework by implementing a PAL module forms one or more groups, and that the PAL modules are programmed so that the Desensitization is done within each group.