CA2236268A1

CA2236268A1 - Method and apparatus for automated image analysis of biological specimens

Info

Publication number: CA2236268A1
Application number: CA002236268A
Authority: CA
Inventors: James E. Ring; James W. Douglass; Thomas J. Riding
Original assignee: Individual
Current assignee: ChromaVision Medical Systems Inc
Priority date: 1995-11-30
Filing date: 1996-11-27
Publication date: 1997-06-05
Also published as: US20080258061A1; WO1997020198A2; US20030228038A1; US20040120562A1; US20080260233A1; DE69627183D1; US6920239B2; US20050185832A1; US20070053569A1; US7783098B2; US7177454B2; DE69627183T2; US7660449B2; US7558415B2; EP0864082A2; US7359536B2; ATE236386T1; US6215892B1; US6553135B1; EP0864082B1

Abstract

A method and apparatus for automated cell analysis of biological specimens automatically scans at a low magnification to acquire images (288) which are analyzed to determine candidate cell objects of interest. The low magnification images are converted from a first color space to a second color space (290). The color space converted image is then low pass filtered (292) and compared to a threshold (294) to remove artifacts and background objects from the candidate object of interest pixels of the color converted image. The candidate object of interest pixels are morphologically processed (296) to group candidate object of interest pixels together into groups which are compared to blob parameters (298) to identify candidate objects of interest which correspond to cells or other structures relevant to medical diagnosis of the biological specimen. The location coordinates of the objects of interest are stored and additional images of the candidate cell objets are acquired at high magnification. The high magnification images are analyzed in the same manner as the low magnification images to confirm the candidate objects of interest which are objects of interest. A high magnification image of each confirmed object of interest is stored for later review and evaluation by a pathologist.

Description

~Y~ ~? ri~nl~8 PCT!l!Sq6/la ME~OD AND APPARATUS FOR AUT~MATE~
IMAGE ANAI YSIS OF BIOLOGICAL SPECIMENS

Backgro~nd of the Tnvention In the field of medical diagnostics including oncology, the detection, identification, quantitation and characterization of cells of interest, such as cancer cells, through testing of biological specimens is an important aspect of diagnosis. Typically, a biological specimen such as bone rnarrow, Iymph nodes, peripheral blood,~rebrospinal f~uid, urine, effusions, fine needle aspirates, peripheral blood scrapings or other materials are ~ arcd by staining the spccim~ll to identify cetls of intercst. One method of cell spccimen pl~pa~dLion is to react a specimen with a specific probe which can be a monoclonal antibody, a polyclonal antiserum, or a nucleic acid which is reactive with a component of the ceils of interest, such as tumor cells. The reaction may be detected using an enzymatic reaction, such as alkaline phosphatase or glucose oxidase or peroxidase to convert a soluble colorless substrate to a colored insoluble precipitate, or by directly conjugating a dye to the probe.
Examination of biological specimens in the past has been perforrned manually by either a lab technician or a pathologist. In the manual method, a slide p,e~a~ed with a biological specimen is viewed at a low magnification under a microscope to visually locate candidate cells of interest.
Those areas of the slide where cells of interest are located are then viewed at a higher magnification to confilm those objects as cells of interest, such as tumor or cancer cells. The manual method is time consuming and prone to error including missing areas of the slide.
Automated cell analysis systems have been developed to improve the speed and accuracy of the testing process. One known interactive system includes a single high power microscope objective for scanning a rack of slidès, portions of which have been previously identified for assay by an ope.dtol. In that system, the operator first scans each slide at a low magnification similar to the manual method and notes the points of interest on the slide for later analysis. The operator then stores the address of the noted location and the associated function in a data file. Once the points of interest have been located and stored by the operator, the slide is then positioned in an automated analysis apparatus which acquires images pf the slide at the marked points and p. .fo~ s an image analysis.

. CA 02236268 1998-04-30 - la -EP-A-O 213 666 discloses a method and apparatus for displaying a single parameter image, in particular a grayscale image. Each pixel of the original image can have a contrast range repre-sented by 8-32 Bits. In order to make this large contrast range visi-ble to the user, it is taught to map a compressed local signal component of the original image into an intensity infor-mation, and a background signal component into a color in~orma-tion of a displayed image.

~'?,~ 8 PCT~!~q6 METHOO AND APPARATUS FOR AUTOMATED
IMAGE ANALYSIS OF BIOLOGICAL SPECIMENS

Back~round of the Invention In th~ field of m~ r 7l r~ eti~ ~ in~ ir~ nnr~ Ey th~ t~ tif~ ntif;r~t jQ~
quantitation and characterization of cells of interest, such as cancer cells, through test~ of biological specimens is an important aspect of diagnosis. Typically, a biological specimsn/such as bone marrow, Iymph nodes, peripheral blood, cerebrospinal fiuid, urine, effusion~fine needle aspirates, peripheral blood scrapings or other materials are ~ ,aled by stainin~/he specimen to identify cells of interest. One method of cell specimen preparation is to re,~ a specimen with a specific probe which can be a monoclonal antibody, a polyclonal antiseru~/or a nucleic acid which is reactive with a component of the cells of interest, such as tumo~ells. The reaction may be detected using an enzymatic reaction, such as alkaline phosphata~r glucose oxidase or peroxidase to convert a soluble colorless substrate to a colored insoluble~écipitate, or by directly conjugating a dye to the probe.
Examination of biological specimens in the p~/has been perforrned manually by either a lab technician or a pathoiogist. In the manual meth~ a slide p.~ alcd with a biological specimen is viewed at a low magnification under a mic~cope to visually locate candidate cells of interest.
Those areas of the slide where cells of in~est are located are then viewed at a higher magnification to confirm those objects as cells of ~rest, such as tumor or cancer cells. The manual method is time consuming and prone to err~ncluding missing areas of the slide.
Automated cell analy~systems have been developed to improve the speed and accuracy of the testing process. O~e/known interactive system includes a single high power microscope objective for scanni~g/a rack of slides, portions of which have been previously identified for assay by an ope.dtor.~that system, the operator first scans each slide at a low magnification similar to the manual ~thod and notes the points of interest on the slide for later analysis. The operator then stores th~ddress of the noted location and the associated function in a data file. Once the points of inte~have been located and stored by the operator, the slide is then positioned in an automated ~Iysis apparatus which acquires images pf the slide at the marlced points and perforrns an image ctn~ly3is.
5~ >
SuTnln~ry of the Inventior A problem with the foregoing autoll~ated system is the continued need for operator input to initially locate cell objects for analysis. Such continued dependence on manual input can lead to W097/20198 PCT~US96/19166 errors including cells of interest being missed. Such errors can be critical especially in assays for so-called rare events, e.g., f~mding one tumor cell in a cell population of one million norrnal cells.
Additionally, manual methods can be extremely time conc~ming and can require a high degree of training to plVpe:l Iy identify and/or quantify cells. This is not only kue for tumor cell detection, but S also for other applications ranging from neutrophil alkaline phosphatase assays, reticulocyte counting and maturation ~c~es~ ent, and others. The associated manual labor leads to a high cost for these procedures in addition to the potential errors that can arise from long, tedious manual exslmin~tions. A need exists, therefore, for an improved automated cell analysis system which can quickly and accurat~ly scan large amounts of biological material on a slide. Accordingly, the 10 present invention provides a method and app~Lus for automated cell analysis which elimin~tes the need for operator input to locate cell objects for analysis.
In accordance with the present invention, a slide ~)I.,pal.,d with a biological specimen and reagent is placed in a slide carrier which preferably holds four slides. The slide carriers are loaded into an input hopper of the automated system. The operator may then enter data identifying the size, 15 shape and location of a scan area on each slide, or, preferably, the system automatically locates a scan area for each slide during slide procescing The opc~tcr then activates the system for slide cçssing. At system activation, a slide carrier is positioned on an X-Y stage of an optical system.
Any bar codes used to identify slides are then read and stored for each slide in a carrier. The entire slide is rapidly scanned at a low magnification, typically lOx. At each location of the scan, a low 20 m~gnification image is acquired and processed to detect candidate objects of interest. Preferably, color, size and shape are used to identify objects of interest. The location of each candidate object of interest is stored.
At the completion of the low level scan for each slide in the carrier on the stage, the optical system is adjusted to a high magnification such as 40x or 60x, and the X-Y stage is positioned to the 25 stored locations for the candidate objects of interest on each slide in the carrier. A high magnification image is acquired for each c~n~ te object of interest and a series of image processing steps are performed to confirm the analysis which was pe.r,ll,lcd at low magnification.
A high magnification image is stored for each confirrned object of interest. These images are then available for retrieval by a pathologist or cytotechnologist to review for final .li~ no~tic evaluation.
30 Having stored the location of each object of interest, a mosaic comprised of the candidate objects of interest for a slide may be ge,.c.dted and stored. The pathologist or cytotechnologist may view the mosaic or may also directly view the slide at the location of an object of interest in the mosaic for further evaluation. The mosaic may be stored on magnetic media for future reference or may be trancmitt~cl to a remote site for review and/or storage. The entire process involved in ex~mining a -2- ~V

WO 97/20198 PCI~/US96/19166 single slide takes on the order of 2-lS minutes depending on scan area size and the number of det.-ct~d candidate objects of interest.
The present invention has utility in the field of oncology for the early detection of minim~l residual disease ("micromet~ct~ces"). Other useful applications include prenatal diagnosis of fetal S cells in maternal blood and in the field of infectious diseases to identify pathogens and viral loads, alkaline phosphatase ~essments, reticulocyte counting, and others.
The procç~sing of images acquired in the automated sc~nning of the present invention preferably includes the steps of transforming the image to a .1irr~ ..L color space; filtering the transfo~ned image with a low pass filter; dynamically thresholding the pixels of the filtered image to 10 suppress background material, performing a morphological function to remove artifacts from the thresholded image; analyzing the thresholded image to deterrnine the presence of one or more regions of connected pixels having the same color; and categorizing every region having a size ~reater than a minimum size as a candidate object of interest.
According to another aspect of the invention, the scan area is ~I~tom~tic~lly determined by 15 sc~nning the slide; acquiring an image at each slide position; analyzing texture inforrnation of each image to detect the edges of the specimen; and storing the locations corresponding to the detected edges to define the scan area.
According to yet another aspect of the invention, automated focusing of the optical system is achieved by initially deterrnining a focal plane from an array of points or locations in the scan area.
20 The derived focal plane enables subsequent rapid automatic focusing in the low power sc~nning operation. The focal plane is det~rrnined by det~rrnining proper focal positions across an array of locations and performing an analysis such as a least squares fit of the array of focal positions to yield a focal plane across the array. Preferably, a focal position at each location is determined by incrementing the position of a Z stage for a fixed number of coarse and fine iterations. ~t each 25 iteration, an image is acquired and a pixel variance or other optical parameter about a pixel mean for the acquired image is c~le~ ted to form a set of variance data. A least squares fit is performed on the variance data according to a known function. The peak value of the least squares fit curve is selected as an estim~te of the best focal position.
In another aspect of the present invention, another focal position method for high 30 m~gnification locates a region of interest centered about a c~n~ te object of interest within a slide which were located during an analysis of the low magnification images. The region of interest is ~ preferably n columns wide, where n is a power of 2. The pixels of this region are then processed using a Fast Fourier Tlan~rolln to generate a spectra of col--pu,-ellL frequencies and corresponding complex m~Enit~lde for each frequency component. Preferably, the complex m~Enit~lde of the VVO97/20198 PCTnJS96/19166 fre~uency components which range from 25% to 75% of the m~ximl]m frequency component are squared and summed to obtain the total power for the region of interest. This process is repeated for other Z positions and the Z position corresponding to the m~imllm total power for the region of interest is selected as the best focal position. This process is preferably used to select a Z position for regions of interest for slides cont~inin~ neutrophils stained with Fast Red to identify alkaline phosphatase in cell cytoplasm and cou~ Lained with hemotoxylin to identify the nucleus of the neutrophil cell. This focal method may be used with other stains and types of biological specimens, as well.
According to still another aspect of the invention, a method and apparatus for automated 10 slide h~n~ling is provided. A slide is mounted onto a slide carrier with a number of other slides side-by-side. The slide carrier is pos;tioned in an input feeder with other slide carriers to facilitate ~utoln~tic analysis of a batch of slides. The slide carrier is loaded onto the X-Y stage of the optical system for the analysis of the slides thereon. Subsequently, the first slide carrier is unloaded into an output feeder after automatic image analysis and the next carrier is automatically loaded.
15 P~rief ~ ;r-n of the 1), ~
The above and other features of the invention including various novel details of construction and combinations of parts will now be more particularly described with ~~ere,~ .,ce to the accompanying drawings and pointed out in the claims. It will be understood that the particular apparatus embodying the invention is shown by way of illustration only and not as a limitation of the 20 invention. The principles and features of this invention may be employed in varied and numerous embodiments without departing from the scope of the invention.
Fig. 1 is a perspective view of an apparatus for automated cell analysis embodying the present invention.
Fig. 2 is a block diagram of the a~ s shown in Fig. 1.
Fig. 3 is a block diagram of the microscope controller of Fig. ~2.
Fig. 4 is a plan view of the a~alalus of Fig. 1 having the housing removed.
Fig. 5 is a side view of a microscope subsystem of the apparatus of Fig. 1.
Fig. 6a is a top view of a slide carrier for use with the apparatus of Fig. 1.
Fig. 6b is a bottom view of the slide carrier of Fig. 6a.
Fig. 7a is a top view of an automated slide h~n~lling subsystem of the a~ dLus of Fig. 1.
Fig. 7b is a partial cross-sectional view of the automated slide h~n~lling subsystem of Fig. 7a taken on line A-A.
Fig. 8 is an end view of the input module of the automated slide h~ntlling subsystem.
Figs. 8a-8d illustrate the input operation of the automatic slide h~nflling subsystem.

wo s7nolss PCT/US96~1gl66 Figs. 9a-9d illustrate the output operation of the ~l-t )m~ted slide h~n-~iing sul~y~L~
Fig. 10 is a flow ~ m of the procedure for automatically determining a scan area.
Fig. 11 shows the scan path on a pl.,l)a~.,d slide in the procedure of Fig. 10.
Fig. 12 illu~tlldles an image of a field acquired in the procedure of Fig. 10.
S Fig. 13A is a flow diagram of a pr~r.~ d proeedure for determining a focal position.
Fig. 13B is a flow diagram of a preferred procedure for de~rrnining a focal position for neuLIul~hils stained with Fast Red and co~ ;.lft~ with hemotoxylin.
Fig. 14 is a flow dia~l~,ll of a procedure for ~ulù~ ;c~lly deL~.-Ilinillg initial foeus.
Fig. 15 shows an array of slide positions for use in the procedure of Fig. 14.
Fig. 16 is a flow ~ r~m of a procedure for auL~ IIlalic focusing at a high magnifieation.
Fig. 17A is a flow diagram of.an overview of the pler~ d process to locate and identify objeets of interest in a stained biological ~ eCinlen on a slide.
Fig. 1 7B is a flow diagram of a procedure for color space conversion.
Fig. I 8 is a flow diagram of a procedure for background suppression via dynamic15 thresholding.
Fig. 19 is a flow diagram of a procedure for molphological prc~cçccin~
Fig. 20 is a flow ~ m of a proeedure for blob analysis.
Fig. 21 is a flow diagram of a procedure for image processing at a high m~gnifiç~tion.
Fig. 22 illustrates a mosaie of cell images produced by the a~alalus~
Fig. 23 is a flow di&~alll of a procedure for e ~ g the number of nucleated cells in a scan area.
Fig. 24 illustrates the apparatus functions available in a user interface of the app&laL~ls.

D-~t~-ilPfl 1 )t~ir. ;~ n of t~- Pr~ f~ d li'nlhn~liment Referring now to the figures, an a~pa.al~s for a-~c....~d cell analysis of biological 25 specimens is generally in~ic~1ed by refe.~nce numeral 10 as shown in pe~ e view in Fig. 1 and in block diagram form in Fig. 2. The apparatus 10 comprises a microscope ~ul~ n 32 housed in a housing 12. The housing 12 includes a slide carrier input hopper 16 and a slide carrier output hopper 18. A door 14 in the housing 12 secures the microscope sub~y~Leln from the external el~vi~ul-l-.fmt A con~ subsystem cûnl~ es a colllp.ll~l 22 having a system processor 23, an 30 image processor 25 and a cnmmllnications modem 29. The computer subsystem further includes a eomputer monitor 26 and an image monitor 27 and other external p~ hclals including storage device 21, track ball device 30, Iceyboard 28 and color printer 35. An external power supply 24 is also shown for powering the system. Viewing oculars 20 of the mi.,l. scopc subsystem project from w o 97nol98 PCT~US96/19166 the housing 12 for opc.dlor viewing. The alJ~JalaLus 10 further includes a CCD eamera 42 for ae~uiring images through the microscope subsystem 32. A microscope eontroller 31 under the eontrol of system proeessor 23 eontrols a number of microseope-subsystem funetions described further in detail. An automatic slide feed meehAni~m 37 in conjunction with X-Y stage 38 provide S ~ torn~tic slide h~n~ling in the al~dlalus 10. An illumination light source 48 projects light onto the X--Y stage 38 which is subsequPntly imaged through the mic-~scvpc subsystem 32 and acquired through CCD eamera 42 for processing in the image pr~,ccssor 25. A Z stage or focus stage 46 under control of the microscope controller 31 provides displacement of the microscope subsystem in the Z
plane for focusing. The mieroscope subsystem 32 further ineludes a rnotorized objeetive turret 44 10 for seleetion of objeetives.
The purpose of the a~*)alalus 10 is for the l-n~ttencled aulOIlldLic sc~nning of ~ ared mi.,roseope slides for the deteetion and eounting of candidate objects of interest such as normal and abnormal cells, e.g., tumor cells. The plefc~l~ d embodiment may be utilized for rare event detection in which there may be only one e~ntli~ te object of interest per several hundred thousand normal 15 eells, e.g., one to five c~n~ te objects of interest per 2 square cf~ e. area of the slide. The â~pdldl~ls 10 autom~tic~lly loeates and eounts candidate objects of interest and ~ es normal cells present in a biologieal ~echllc.l on the basis of color, size and shape eharact~.;sLics. A number of stains are used to p..,~.. .ILially stain e~n~ te objects of interest and normal eells dirr..~ eolors so that sueh eells ean be distinguished from eaeh other.
As noted in the baekground of the invention, a biological speeimen may be ~ Jdl. d with a reagent to obtain a colored insoluble precipiLdl~. The a~pdlalus of the present invention is used to deteet this ~l~ei~ L~; as a c m~ te object of interest.
During operation of the appdlal.ls 10, a pathologist or laboratory technician mounts ~lel,a~d slides onto slide carriers. A slide carrier 60 is illustrated in Fig. 8 and will be deseribed further below. Each slide earrier holds up to 4 slides. Up to 25 slide earriers are then loaded - into input hopper 16. The Op~,.dtOI can specify the size, shape and loeation of the area to be seanned or alternatively, the system ean automatieally loeate this area. The op.. dLor then co.. ~ the system to begin aulc~ll.aled sc~nninF~ of the slides through a graphical user il,l .~dce. Unattended sc~nning begins with the automatie loading of the first earrier and slide onto the preeision motorized X-Y
30 stage 38. A bar eode label affixed to the slide is read by a bar eode reader 33 during this loading operation. Eaeh slide is then seanned at a user seleeted low mieroseope magnifieation, for example, lOx, to identify c~n~ te eells based on their eolor, size and shape eharaeteristics. The X-Y
locations of c nrli~ tlo eells are stored until sc~nning is completed.

W O 97/20198 P ~ nUS96~9166 After the low magnification sc~nning is completed, the al~paldlus ~ o~ cally returns to each ç~n~1id~te cell, reimages and refocuses at a higher m~gnific~tinn such as 40x and p~,lrulllls further analysis to confirm the cell candidate. The a~a,al.ls stores an image of the cell for later review by a pathologist. All results and images can be stored to a storage device 21 such as a ' 5 removable hard drive or DAT tape or h-~n~mitted to a remote site for review or storage. The stored images for each slide can be viewed in a mosaic of images for further review. In addition, the pathologist or operator can also directly view a detected cell through the microscope using the included oculars 20 or on image monitor 27.
Having described the overall operation of the a~dlaLus 10 from a high level, the further 10 details of the apl)aldlus will now be described. Referring to Fig. 3, the mi.r..scopc controller 31 is shown in more detail. The mic-uscopc controller 31 includes a number of s~,b~y~ ls comlecle;d through a system bus. A system pl~Jcci,~of 102 controls these subsystems and is controlled by the ~)paldlUS system processor 23 through an RS 232 controller 110. The system processor 102 controls a set of motor - control subsystems 114 through 124 which control the input and output feeder, the motorized turret 44, the ~--Y stage 3g, and the Z stage 46 (Fig. 2). A histogram processor 108 ~cei~s input from CCD camera 42 for cu,..~ variance data during the focusing operation described further herein.
The system processor 102 further controls an illu~ lion controller 106 for control of substage illll...;..A~;on 48. The light output from the halogen light bulb which supplies ill--min~tion for the system can vary over time due to bulb aging, r.h~n~es in optical alignment, and other factors.
In addition, slides which have been "over stained" can reduce the camera e~o~ule to an unacc~Ldble level. In order to coll,pe~ L~ for these effects, the illllmin~tion controller 106 is included. This controller is used in conjunction with light control software to coln~e.lsdle for the variations in light level. The light control software samples the output from the camera at interva1s (such as between loading of slide carriers), and comm~ntle the controller to adjust the light level to the desired levels. In this way, light control is automatic and Lla~ ll to the user and adds no additional time to system operation.
The system processor 23 is preferably com~.ised of dual parallel Intel ~ lLiulll 90 MHz devices. The image processor 25 is preferably a Matrox ~m~ing Series 640 model. The microscope controller system processor 102 is an Advanced Micro Devices AMD29K device.
Referring now to Figs. 4 and 5, further detail of the ~ dlus 10 is shown. Fig. 4 shows a plan view of the dpp~UdtUS 10 with the housing 12 removed. A portion of the ~lltom~tic slide feed mecllallislll 37 is shown to the left of the mic.osco,oe subsystem 32 and includes slide carrier W O 97/20198 PCTnUS96/19166 unloading assembly 34 and llnlo~iing plalro..,l 36 which in conju...;lion with slide carrier output hopper 18 function to receive slide carriers which have been analyzed.
Vibration isolation mounts 40, shown in further detail in Fig. 5, are provided to isolate the microscope subsystem 32 from ",ecllAnic~l shock and vibration that can occur in a typical laboratory 5 ~nvirullc.-l. In ~l"ition to ~~t~ sources of vibration, the high speed operation of the X-Y stage 38 can induce vibration into the microscope ,~lb~y~ 32. Such sources of vibration can be isolated from the electro-optical subsystems to avoid any un-~e~irahle effects on image quality. The isolation mounts 40 comprise a spring 40a and piston 40b ,ubln~ ;ed in a high viscosity silicon gel which is enclosed in an elastomer membrane bonded to a casing to achieve damping faetors on the order of 17to20%.
The automatic slide h~nllling feature of the present invention will now be described. The autom~ted slide h~ndling sub.,y,L~n. op~ tes on a single slide carrier at a time. A slide carrier 60 is shown in Figs. 6a and 6b which provide a top view and a bottom view ~G ,~,c.,Livt;ly. The slide carrier 60 includes up to four slides 70 mounted with adhesive tape 62. The carrier 60 includes ears 64 for h~nging the carrier in the output hopper 18. An undercut 66 and pitch rack 68 are formed at the top edge of the slide carrier 60 for mechanical h~n-llin~ of the slide carrier. A keyway cutout 65 is formed in one side of the carrier 60 to facilitate carrier ~ nment A ~l~p,llt;d slide 72 mounted on the slide carrier 60 includes a sample area 72a and a bar code label area 72b.
Fig. 7a provides a top view of the slide h~ndling subsystem which comprises a slide input module 15, a slide output module 17 andX-Ystage drive belt 50. Fig. 7b provides a partial cross-sectional view taken along line A--A of Fig. 7a.
The slide input module 15 co.,.~,.;ses a slide carrier input hopper 16, loading platform 52 and slide carrier loading ~ cse~..hly 54. The input hopper 16 lecc;ves a series of slide carriers 60 (Figs.
6a and 6b) in a stack on loading platform 52. A guide key 57 protrudes from a side of the input hopper 16 to which the keyway cutout 65 (Fig. 6a) of the carrier is fit to achieve proper ~ nment The input module 15 further includes a revolving in~ieYing cam 56 and a switch 90 mo~mted in the loading platform 52, the operation of which is described further below. The carrier loading sf-..bly 54 comprises an infeed drive belt 59 driven by a motor 86. The infeed drive belt 59 includes a pusher tab 58 for pushing the slide carrier horizontally toward the X-Y stage 38 when the 3û belt is driven. A homing switch 95 senses the pusher tab 58 during a revolution of the belt 59.
Referring specifically to Fig. 7a, the X-Y stage 38 is shown with x position and y position motors 96 and 97 les~e~ ely which are controlled by the mic,vscopc controller 31 (Fig. 3) and are not considered part of the slide h~n~ling sul,~y~L~lll. The X--Y stage 38 further includes an al)e.~
55 for allowing illl-min~tion to reach the slide carrier. A switch 91 is mounted ~dj~l~ent the apf ,Lure -CA 02236268 l998-04-30 W ~ 97~198 PCT~JS96/19~66 55 for sensing contact with the carrier and thereupon activating a motor 87 to drive stage drive belt 50 (Fig. 7b). The drive belt 50 is a double sided timing belt having teeth for ene~ging pitch rack 68 of the carrier 60 (Fig. 6b).
The slide output module 17 includes slide carrier output hopper 18, unloading platform 36 S and slide carrier unloading sllb~e~embly 34. The unloading ~ s~hly 34 comprises a motor 89 for rotating the unloading platform 36 about shaft 98 during an unloading operation described further below. An outfeed gear 93 driven by motor 88 rotatably engages the pitch rack 68 of the carrier 60 (Fig. 6b) to l~ ~ull the carrier to a rest position against switch 92. A springloaded hold-down mec~ni~m holds the carrier in place on the unloading platform 36.
The slide handling operation will now be described. Referring to Fig. 8, a series of slide carriers 60 are shown stacked in input hopper 16 with the top edges 60a ~li~e-l As the slide h~n~11ing operation begins, the inclexinE cam 56 driven by motor 85 advances one revolution to allow only one slide carrier to drop to the bottom of the hopper 16 and onto the loading platform 52.
Figs. 8a--8d show the cam action in more detail. The cam 56 includes a hub 56a to which are mounted upper and lower leaves 56b and 56c respectively. The leaves 56b, 56c are semicircular projections oppositely positioned and spaced apart vertically. In a first position shown in Fig. 8a, the upper leaf 56b :iUppOl~ the bottom carrier at the undercut portion 66. At a position of the cam 56 rotated 180~, shown in Fig. 8b, the upper leaf 56b no longer supports the carrier and instead the carrier has dropped slightly and is supported by the lower leaf 56c. Fig. 8c shows the position of the cam 56 rotated 270~ wl.~,le;ll the upper leaf 56b has rotated sufficiently to begin to engage the undercut 66 of the next slide carrier while the opposite facing lower leaf 56c still supports the bottom carrier. After a full rotation of 360~ as shown in Fig. 8d, the lower leaf 56c has rotated opposite the carrier stack and no longer supports the bottom carrier which now rests on the loading platform 52. At the same position, the upper leaf 56b :iU~OltS the next carrier for l~,pe~.tillg the cycle.
Referring again to Figs. 7a and 7b, when the carrier drops to the loading platform 52, the contact closes switch 90 which activates motors 86 and 87. Motor 86 drives the infeed drive belt S9 until the pusher tab 58 makes contact with the carrier and pushes the carrier onto the X-Y stage drive belt 50. The stage drive belt 50 advances the carrier until contact is made with switch 91, the closing of which begins the slide sc~nnin~ process descl;l,ed further herein. Upon completion of the sc.annin~ process, the X-Y stage 38 moves to an unload position and motors 87 and 88 are activated to transport the carrier to the unloading platform 36 using stage drive belt S0. As noted, motor 88 drives outfeed gear 93 to engage the carrier pitch rack 68 of the carrier 60 (~ig. 6b) until switch 92 is c~ nt~cted Closing switch 92 activates motor 89 to rotate the unloading platform 36.

g W O 97/20198 PCTAUS96/19166The unloading operation is shown in more detail in end views of the output module 17 (Figs.
9a-9d). In Fig. 9a, the unlo~-linE platform 36 is shown in a horizontal position ~ ~o~ling a slide carrier 60. The hold-down ,l,ec~ m 9~ secures the carrier 60 at one end. Fig. 9b shows the output module 17 after motor 89 has rotated the unloading platform 36 to a vertical position, at which point the spring loaded hold-down l.,ecl-A~ m 94 releases the slide carrier 60 into the output hopper 18. The carrier 60 is supported in the output hopper 18 by means of ears 64 (Figs. 6a and 6b). Fig. 9c shows the unloading platform 36 being rotated back towards the horizontal position. As the platform 36 rotates upward, it contacts the d~,~,osil~d carrier 60 and the upward movement pushes the carrier toward the front of the output hopper 18. Fig. 9d shows the unloading platform 36 at its 10 original horizontal position after having output a series of slide carriers 60 to the output hopper 18.
Having described the overall system and the automated slide h~n~lling feature, the aspects of the a~pdlaLùs 10 relating to sC~rlning~ focusing and image processing will now be described in further detail.
In some cases, an op~al(JI will know ahead of time where the scan area of interest is on the 1~ slide. Conventional ~ Jalalion of slides for ~o~ min~tion provides repeatable and known placement of the sample on the slide. The OPG-~aLOr can the.erole instruct the system to always scan the same area at the same location of every slide which is ~.epalGd in this fashion. But there are other times in which the area of interest is not known, for example, where slides are ~ol~a.Gd m~n~ ly with a known smear technique. One feature of the invention a~llolo~;c~lly determines the scan area using a 20 texture analysis process.
Fig. 10 is a flow diagram that describes the processing associated with the ~utom~tiC location of a scan area. As shown in this figure, the basic method is to pre-scan the entire slide area to determine texture fedlu~Gs that indicate the p.es...ce of a smear and to discriminate these areas from dirt and other artifacts.
At each location of this raster scan, an image such as in Fig. 12 is ac~luhGd and analyzed for texture il-rc,..l.aLion at steps 204 and 206. Since it is desired to locate the edges of the smear sample within a given image, texture analyses are cond~lcted over areas called windows 78, which are smaller than the entire image as shown in Fig. 12. The process iterates the scan across the slide at steps 208, 210, 212 and 214.
In the interest of speed, the texture analysis process is pc~ru.~,led at a lower magnification, ~cr~._bly at a 4x objective. One reason to operate at low m~gnific~tion is to image the largest slide area at any one time. Since cells do not yet need to be resolved at this stage of the overall image analysis, the 4x m~ ;on is preferred. On a typical slide, as shown in Fig. l l, a po~tion 72b of W Og7/20198 PCT~US96/19166 the end of the slide 72 is reserved for labeling with iclentifi~tion h~r~llllaLion. Excepting this label area, the entire slide is sc~nned in a raster scan fashion 76 to yield a number of adjacent images.
Texture values for each window include the pixel variance over a window, the difference beLvY~ ell the largest and ~m~ st pixel value within a window, and other indicators. The ~,c.sence of 5 a smear raises the texture values coll,palcd with a blank area.
One problem with a smear from the standpoint of dcte- ...i..i..g its location is its non-uniform thickness and texture. For example, the smear is likely to be relatively thin at the edges and thicker towards the middle due to the nature of the smearing process. To accommodate for the non-uniformity, texture analysis provides a texture value for each analyzed area. The texture value 10 tends to gradually rise as the scan proceeds across a smear from a thin area to a thick area, reaches a peak, and then falls off again to a lower value as a thin area at the edge is reached. The problem is then to decide from the series of texture values the be~;..,.il-g and ending, or the edges, of the smear.
The texture values are fit to a square wave waver~,lm since the texture data does not have sharp beginning~ and enlling.e.
After con~t~ting this sc~nning and texture evaluation operation, one must <~ete~nine which areas of elevated texture values lcpl~elll the desired smear 74, and which lGpre~e~lt undesired artifacts. This is accomplished by fitting a step function, on a line by line basis to the texture values in step 216. This function, which resembles a single square wave across the smear with a begimlillg at one edge, and end at the other edge, and an amplitude provides the means for discrimination. The 20 amplitude of the best-fit step function is utilized to determine whe~ smear or dirt is present since relatively high values indicate smear. If it is decided that smear is present, the beginning and ending cooldillales of this pattern are noted until all lines have been ploce~sed, and the smear sample area defined at 218.
After an initial focusing operation described further herein, the scan area of interest is 25 sc~nn~d to acquire images for image analysis. The ~ r.,.led method of operation is to initially p~,~r~l,ll a complete scan of the slide at low m~gnifi~tion to identify and locate c~n~ te objects of interest, followed by further image analysis of the c~n~ te objects of interest at high m~nificsltion in order to confirm the objects as cells. An alternate method of operation is to perform high magnification image analysis of each candidate object of interest immediately after the object has 30 been identified at low m~nifcation. The low m~gnific~tion sc~nnin~ then resumes, searching for additional candidate objects of interest. Since it takes on the order of a few seconds to change objectives, this alternate method of operation would take longer to complete.
The c~laL~r can pre-select a magnification level to be used for the sc~nning operation. A
low magnification using a 10x objective i5~ fe.l~d for the scanning operation since a larger area W O 97/20198 PCT~USg6/19166 can be initially analyzed for each acquired scan image. The overall detection process for a cell includes a combination of decisions made at both low (lOx) and high m~gnific~tion (40x) levels.
Decision making at the lOx magnification level is broader in scope, i.e., objects that loosely flt the relevant color, size and shape characteristics are identified at the lOx level. Analysis at the 40x S magnification level then proceeds to reflne the decision making and confirm objects as likely cells or candidate objects of interest. For exarnple, at the 40x level it is not u~co~"~"on to find that some objects that were identified at lOx are artifacts which the analysis process will then reiect. In addition, closely packed objects of interest a~peali"g at lOx are s~l)alal~d at the 40x level.
In a situation where a cell straddles or overlaps ~dj~cent image fields, image analysis of the 10 individual adjacent image fields could result in the cell being rejected or un~tecte-1 To avoid missing such cells, the sc~nning operation cv~ es by ovella~hlg ad3~cent image ~lelds in both the x and y directions. An overlap amount greater than half the ~ met~r of an average cell is pl~ifell.,d. In the ~,.erell~,d embodiment, the overlap is specified as a p~ ge of the image field in the x and y directions.
The time to complete an image analysis can vary depending upon the size of the scan area and the number of c~n-iid~te cells, or objects of interest identified. For one G"an"ulc, in the plcfcl.~d embodiment, a complete image analysis of a scan area of two square centim~ot~rs in which 50 objects of interest are confirmed can be p~,lr~ lcd in about 12 to 15 min~lt~e This example includes not only focusing, sc~nning and image analysis but also the saving of 40x images as a mosaic on hard 20 drive 21 (Fi{~. 2).
Consider the utility of the present invention in a "rare event" application where there may be one, two or a very small number of cells of interest located somewhere on the slide. To illustrate the nature of the problem by analogy, if one were to scale a slide to the size of a football field, a tumor cell, for example, would be about the size of a bottle cap. The problem is then to rapidly search the 25 football field and find the very small number of bottle caps and have a high c~,l~illly that none have been missed.
However the scan area is defined, an initial focusing operation must be perforlned on each slide prior to sc~ -g This is required since slides differ, in general, in their pl~r~m~nt in a carrier.
These differences include slight (but signifir~nt) variations of tilt of the slide in its carrier. Since 30 each slide must remain in focus during sc~nning, the degree of tilt of each slide must be determined.
This is accomplished with an initial focusing operation that del~l..,i,les the exact degree of tilt, so that focus can be ~.. ;..u ;.-cd autom~tir~lly during SC~nning The initial focusing operation and other focusing operations to be described later utilize a focusing method based on ~ Jce;.~;ng of images acquired by the system. This method was chose W O 97~0198 PCTAUS96/19166 for its simplicity over other methods including use of IR beams reflected from the slide surface and use of mechanical gauges. These other methods also would not function ~r~/~clly when the specimen is l~lolecled with a coverglass. The p~c~llGd method results in lower system cost and improved reliability since no additional parts need be included to p.,.rOllll focusing.
Fig. 13A provides a flow diagram describing the "focus point" procedure. l~e basic method relies on the fact that the pixel value variance (or standard deviation) taken about the pixel value mean is m~ximnnn at best focus. A "brute-force" method could simply step through focus, using the cc~ uL~,. controlled Z, or focus stage, calculate the pixel variance at each step, and return to the focus position providing the m;-X;,..l--.. v~;ancc. Such a method would be too time con~--ming Therefore, additional features were added as shown in Fig. 13A.
These features include the determin~tion of pixel variance at a relatively coarse number of focal positions, and then the fitting of a curve to the data to provide a faster means of determining optimal focus. This basic process is applied in two steps, coarse and fine.
During the coarse step at 220-230, the Z stage is stepped over a user-specified range of focus po~ition~, with step sizes that are also user-specified. It has been found that for coarse focusing, these data are a close fit to a G~ c;~n function. Therefore, this initial set of variance versus focus position data are least-squares fit to a G~c~i~n function at 228. The location of the peak of this C'J~ncc;~n curve deterrninec the initial or coarse c il ;. . .~lt; of focus position for input to step 232.
Following this, a second stepping operation 232-242 is p~,.rc,ll,led ~Itili7in~ smaller steps over a smaller focus range centered on the coarse focus position. Experience in~lic~tec that data taken over this smaller range are generally best fit by a second order polynomial. Once this least squares fit is performed at 240, the peak of the second order curve provides the fine focus position at 244 Fig 14 illustrates a procedure for how this focusing method is utilized to (ietçrrnine the orientation of a slide in its carrier. As shown, focus positions are determined, as described above, for a 3 x 3 grid of points centered on the scan area at 264. Should one or more of these points lie outside the scan area, the method senses at 266 this by virtue of low values of pixel variance. In this case, ~ddition~l points are selected closer to the center of the scan area. Fig. 15 shows the initial array of points 80 and new point 82 selected closer to the center. Once this array of focus positions is determined at 268, a least squares plane is fit to this data at 270. Focus points Iying too far above or below this best-fit plane are discarded at 272 (such as can occur from a dirl~ cover glass over the scan area), and the data is then refit. This plane at 274 then provides the desired Z position information for rn~int~ining focus during sc~nning After determination of the best-fit focus plane, the scan area is sc~nnPd in an X raster scan over the scan area as described earlier. During sc~nning the X stage is positioned to the starting CA 02236268 l998-04-30 W O 97/20198 PCT~US96/19166 point of the scan area, the focus (Z) stage is positioned to the best fit focus plane, an image is acquired and processed as described later, and this process is repeated for all points over the scan area. In this way, focus is m~int~ined autom~tic~lly without the need for time-con~-nning refocusing at points during sçSlnning S Prior to conf~ tion of cell objects at a 40x or 60x level, a refocusing operation is con-lucted since the use of this higher m~ific~tion requires more precise focus than the best-fit plane provides. Fig. 16 provides the flow diagram for this process. As may be seen, this process is similar to the fine focus method described earlier in that the object is to m~imi7P the image pixel iancc. This is accomplished by stepping through a range of focus positions with the Z stage at 10 276, 278, calculating the image v~,;ancc at each position at 278, fitting a second order polynomial to these data at 282, and calculating the peak of this curve to yield an estim~te of the best focus position at 284, 286. This final focusing step differs from previous ones in that the focus range and focus step sizes are smaller since this magnification ~e.~ui,cs focus settings to within 0.5 micron or better.
It should be noted that for some combinations of cell st~inin~ characteristics, improved focus can be obtained by numerically selecting the focus position that provides the largest variance, as opposed to selecting the peak of the polynomial. In such cases, the polynomial is used to provide an estim~te of best focus, and a final step selects the actual Z position giving highest pixel variance. It should also be noted that if at any time during the focusing process at 40x or 60x the parameters 20 inf1ic~te that the focus position is in~-lequ~te the system automatically reverts to a coarse focusing process as described above with rc;r~,.cncc to Fig. 13A. This ensures that variations in specil.le thickness can be accommodated in an ~xpe~itious manner.
For some biological specimens and stains, the focusing methods ~i~cusse~i above do not provide optimal focused results. For example, certain white blood cells known as neutrophils may 25 be stained with Fast Red, a conllllollly known stain, to identify alkaline phQsl~h~lztce in the cy~opla~ll, of the cells. To further identify these cells and the material within them, the specimen may be counl~ cd with hemotoxylin to identify the nucleus of the cells. In cells so treated, the cytoplasm bearing alkaline phosph~t~e becomes a shade of red proportionate to the amount of alkaline phosph~t~e in the cytoplasm and the nucleus becomes blue. However, where the 30 cytoplasm and nucleus overlap, the cell appears purple. These color conlbill~lions appear to preclude the finding of a focused Z position using the focus processes tli~c~-c~ed above.
In an effort to find a best focal position at high magnification, a focus method, such as the one shown in Fig. 13B, may be used. That method begins by selecting a pixel near the center of a c~ntlic~te object of interest (Block 248) and defining a region of interest centered about the selected WO 97/20198 PCT~US96/19166 pixel (Block 250). Preferably, the width of the region of interest is a number of columns which is a power of 2. This width l),efe,~llce arises from subsequent processing of the region of interest preferably using a one dimensional Fast Fourier Transform (FFT) technique. As is well known within, the art, processing columns of pixel values using the FFT technique is facilitated by making 5 the number of columns to be processed a power of two. While the height of the region of interest is also a power of two in the ~l~rcll~d embodiment, it need not be unless a two dimensional FFT
tecllni~ e is used to process the region of interest.
After the region of interest is selected, the columns of pixel values are processed using the ~,~r~ d one dimensional FFT to determine a spectra of frequency components fo} the region of 10 interest (Block 252). The frequency spectra ranges from DC to some highest frequency component.
For each frequency coln~o~ , a c~lllpl m~nitll~P is co~ ul~d. Preferably, the colll~
m~nit~ for the frequency con,~uolle.-l~ which range from al,,uloxhl~ately 25% of the highest component to a~l)loxilndt~ly 75% of the highest component are squared and summed to detçrrnine the total power for the region of interest (Block 254). All~lualively, the region of interest may be 15 p}ocessed with a smoothing window, such as a Hanning window, to reduce the spurious high frequency components generated by the FFT processing of the pixel values in the region of interest.
Such preproces~in~ of the region of interest permits all complex magnitude over the complete frequency range to be squared and summed. After the power for a region has been coll.~.uled and stored (Block ~56), a new focal position is selected, focus adjusted ~Blocks 258, 260), and the 20 process repeated. After each focal position has been evaluated, the one having the greatest power factor is selected as the one best in focus (Block 262).
The following describes the image proces;,illg methods which are utilized to decide whether a candidate object of interest such as a stained tumor cell is present in a given image, or field, during the sc~nnin~ process. C~ndi~te objects of interest which are ~letected during sc~....;.,~ are reimaged 25 at higher ~40x or 60x) m~gnific~tion, the decision cc~l.ri....~d, and a region of interest for this cell saved for later review by the pathologist.
The image processing includes color space cull~ ion, low pass filtering, background .ession, artifact suppression, morphological l luces~ g~ and blob analysis. One or more of these steps can optionally be elimin~ted The ~ ul is provided with an option to configure the system 30 to ~ ~,.rc,~ any or all of these steps and whether to perform certain steps more than once or several times in a row. It should also be noted that the sequence of steps may be varied and thereby - optimi7ed for specific reagents or reagent combinations; however, the sequence described herein is r~ d. It should be noted that the image processing steps of low pass filtering, thresholding, morphological p..)cc~ g, and blob analysis are generally known image processing building blocks.

w o 97nO198 PCTAJS96/19166 An overview of the preferred process is shown in Fig. 17A. The p,~r~ d process for identifying and locating c~ndi~1~te objects of interest in a stained biological specimen on a slide begins with an acquisition of images obtained by sc~nninp: the slide at low ma~l,if~ccalion (Block 288). Each image is then converted from a first color space to a second color space (Block 290) and the color converted image is low pass filtered (Block 292). The pixels of the low pass filtered image are then co,-.pal~d to a threshold (Block 2g4) and, preferably, those pixels having a value equal to or greater than the threshold are identified as candidate object of interest pixels and those less than the threshold are dcl~ lh-ed to be artifact or background pixels. The c~n~ te object of interest pixels are then morphologically processed to identify groups of c~nrl;~te object of interest pixels as 10 c~n~ te objects of interest (Block 296). These c~nr~ te objects of interest are then co",~ d to blob analysis parameters (Block 298) to further dirr~ e c~n~ te objects of interest from objects which do not conro"" to the blob analysis ~u~lalll~ and, thus, do not warrant further procec~in~ The location of the c~n~ te objects of interest may be stored prior to confirm~tion at high magnification. The process continues by d~ whether the ç~n~ te objects of interest 15 have been confirmed (Block 300). If they have not been conrll".cd, the optical system is set to high m~gnific~tion (Block 302) and images of the slide at the locations co,l~onding to the candidate objects of interest i~l~ntified in the low magnification images are acquired (Block 288). These images are then color converted (Block 290), low pass filtered (Block 292), conlpa~td to a threshold (Block 294), morphologically processed (Block 296), and co~ aled to blob analysis parameters 20 (Block 298) to confirm which c~n~l;d?~te objects of interest located from the low magnification images are objects of interest. The coordil,~l~s of the objects of interest are then stored for future reference (Block 303).
Neural net processing s.;l,ell,es were not considered for the ~)le~"~id embodiment for several reasons. Firstly, the ~ d embodiment is c,~Lill~i~~d for "rarc-evclll" detection, although it is not 25 limited to this case. Since neural nets must be trained on what to look for, sometimes several thousands Of examples must be plGi,~llL~d to the neural net for this training. This is impractical for a rare-event application. Secondly, neural net processing can be slower than "~ tic~ methods, sometimes by large factors. Th~,lerole, neural nets were not deemed ~ pl;ate for this application, ~lthough certain features of the invention may be advantageously applied to neural n~ Lwolh systems.
In general, the candidate objects of interest, such as tumor cells, are ~letect~l based on a combination of characteristics, including size, shape, and color. The chain of decision making based on these cha,~cl~,;stics preferably begins with a color space conversion process. The CCD camera coupled to the microscope subsystem outputs a color image comprising a makix of 640 x 480 pixels.
Each pixel comprises red, green and blue (RGB) signal values.

W O 97/20198 PCT~US96/19I66 It is desi,~.blc to transform the matrix of RGB values to a dir~e..l color space because the dirr~ ce between c~n~ te objects of interest and their background, such as tumor and normal cells, may be det~ rd from their ..~ .~e.~live colors. Speci~lells are generally stained with one or more industry ~ dald stains (e.g., ~AB, New Fuchsin, AEC) which are "reddish" in color.
5 Candidate objects of interest retain more of the stain and thus appear red while normal cells remain lln~t~ined. The specimens may also be cO~ f ~l~;..ed with hemAt ~c lin so the nuclei of normal cells or cells not cont~ining an object of interest appear blue. In addition to these objects, dirt and debris can appear as black, gray, or can also be lightly stained red or blue depending on the staining p~vcedu~s lltili7~-1 The residual plasma or other fluids also present on a smear may also possess 10 some color.
In the color conversion operation, a ratio of two of the RGB signal values is formed to provide a means for di~.;...;..~l;..g color in~.l..alion. With three signal values for each pixel, nine dirr~ ratios can be formed:
, R/G, R/B, G/G, G/B, C~/R, B/B, BM, B/R
15 The optimal ratio to select depends upon the range of color information expected in the slide specimen. As noted above, typical stains used for dcte.~ g c~n~ te objects of interest such as tumor cells are predominantly red, as opposed to predominantly green or blue. Thus, the pixels of a cell of interest which has been stained contain a red cc.~l,poncnt which is larger than either the green or blue colllpon~ i. A ratio of red divided by blue (R/B) provides a value which is greater than one 20 for tumor cells but is approximately one for any clear or white areas on the slide. Since the rem~ining cells, i.e., normal cells, typically are stained blue, the R/B ratio for pixels of these latter cells yields values of less than one. The R/B ratio is ple~...~d for clearly sct~ala~ g the color hlrol.lla~ion typical in these appli~ ti~n~
Fig. 17B illustrates the flow diagram by which this conversion is performed. In the interest 25 of processing speed, the collv~l~ion is implemented with a look up table. The use of a look up table for color conversion accomplishes three functions: 1) p.,lrolllling a division operation; 2) scaling the result for processing as an image having pixel values ranging from 0 to 255; and 3) defining objects which have low pixel values in each color band ~R,G,B) as "black" to avoid infinite ratios (i.e., dividing by zero). These "black" objects are typically staining artifacts or can be edges of bubbles ~ 30 caused by pasting a coverglass over the ~tet~illlen.
Once the look up table is built at 304 for the specific color ratio (i.e., choices of tumor and - nucleated cell stains), each pixel in the original RGB image is coll~ Led at 308 to produce the output. Since it is of interest to scpalal~ the red stained tumor cells from blue stained norrnal ones, the ratio of color values is then scaled by a user specified factor. As an example, for a factor of 128 W O 97/20198 PCT~US96/19166and the ratio of (red pixel value)/(blue pixel value~, clear areas on the slide would have a ratio of I
scaled by 128 for a final X value of 128. Pixels which lie in red stained tumor cells would have X
value greater than 128, while blue stained nuclei of normal cells would have value less than 128. In this way, the desired objects of interest can be numerically dis.,~ ed The resulting 640 x 480 5 pixel matrix, referred to as the X-image, is a gray scale image having values ranging from 0 to 255.
Other methods exist for discrimin~ting color information. One classical method convelL~ the RGB color information into another color space, such as HSI (hue, saturation, illk;n~il;y) space. In such a space, distinctly dirr,~.,l hues such as red, blue, green, yellow, may be readily selJalat~d. In addition, relatively lightly stained objects may be ~i~tinglliched from more intensely stained ones by 10 virtue of differing saturations. However, converting from RGB space to HSI space requires more complex co~u~,ul~lion. Conversion to a color ratio is faster; for e~alll,lJlc, a full image can be converted by the ratio teehni,lue of the present invention in about 30 ms while an HSI conversion can take several seconds.
In yet another approach, one could obtain color inforrnation by taking a single color channel 15 from the camera. As an example, consider a blue channel, in which objects that are red are relatively dark. Objects which are blue, or white, are relatively light in the blue channel. In principle, one could take a single color channel, and simply set a threshold wherein everything darker than some threshold is ca~;uli~ed as a ç~n~irl~te object of interest, for example, a tumor cell, because it is red and hence dark in the channel being reviewed. However, one problem with the single channel 20 a~ ach occurs where illllmin~tion is not uniform. Non-ulliro-llliLy of illumination results in non-uniformity across the pixel values in any color channel, for example, tending to peak in the middle of the image and dropping off at the edges where the illumination falls off. Performing thresholding on this non-uniforrn color information runs into problems, as the edges sometimes fall below the threshold, and thel~er~ it becomes more difficult to pick the ~pl~l.l;ate threshold level.
25 However, with the ratio technique, if the values of the red channel fall off from center to edge, then the values of the blue channel also fall off center to edge, resulting in a uniforrn ratio. Thus, the ratio technique is more immnn~ to illumin~tion non-uniformities.
As previously described, the color conversion scheme is relatively h~sensiLive to changes in color b~i~nce~ i.e., the relative outputs of the red, green, and blue channels. However, some control 30 is l-~ces~. y to avoid camera saturation, or in~dequ~te exposures in any one of the color bands. This color balancing is performed auLo-~ ic~lly by lltil j7.illg a calibration slide consisting of a clear area, and a "dark" area having a known optical tr~n~mi~sion or density. The system obtains images from the clear and "dark" areas, calculates "white" and "black" ad~ tnn~nt~ for the image ~ ,ce3sor 25, and thereby provides correct color /c~l~nce W O 97~98 Pcr~usg6ngl66 In addition to the color balance control, certain mechanical ~ nm~?nte are automated in this process. The center point in the field of view for the various mi~;,oscop~ objectives as measured on the slide can vary by several (or several tens of) microns. This is the result of slight variations in posi~ion of the microscope objectives 44a as detçrmined by the turret 44 (Fig. 4), small variations in '5 ~lignm.ont of the objectives with respect to the system optical axis, and other factors. Since it is desired that each microscope objective be centered at the same point, these mechanical offsets must be measured and automatically colllpe~ sd This is accomplished by im~in~ a test slide which contains a recognizable feature or mark.
An image of this pattern is obtained by the system with a given objective, and the position of the 10 mark detersnined. The system then rotates the turret to the next lens objective, obtains an image of the test object, and its position is redetermined. Apparent changes in position of the test snark are recorded for this objective. This process is continued for all objectives..
Once these spatial offsets have been deterrnined, they are autom~tir~lly cG~ c~ d for by moving the stage 38 by an equal (but opposite) amount of offset during changes in objective. In this 15 way, as di~-~,n~ Iens objectives are selected, there is no a~le.-l shift in center point or area viewed.
A low pass filtering process precedes thresholding. An objective of thresholding is to obtain a pixel image matrix having only candidate objects of interest, such as tumor cells above a threshold level and everything e}se below it. However, an act sal acquired image will contain noise. The noise 20 can ta ce several forrns, including white noise and artifacts. The microscope slide can have small fr~gm~nt~ of debris that pick up color in the staining process and these are known as artifacts. These allir~ are generally small and scattered areas, on the order of a few pixels, which are above the threshold. The purpose of low pass fsltering is to ess~nti~lly blur or smear the entire color converted image. The low pass filtering process will smear artifacts more than larger objects of interest. such 25 as tumor cells and thereby etimin~te or reduce the number of artifacts that pass the thresholding process. l'he result is a cleaner thresholded image dow.l~L.c~..n.
In the low pass filter process, a 3 x 3 matrix of coefficients is applied to each pixel in the 640 x 480 x-image. ~ pserc~lcd coefficient matrix is as follows:

W O 97/20198 PCT~US96/19166At each pixel location, a 3 x 3 matrix co~ .l;sil~g the pixel of interest and its neighbors is multiplied by the coefficient matrix and sllmmed to yield a single value for the pixel of interest. The output of this spatial convolution process is again a 640 x 480 matrix.
As an example, consider a case where the center pixel and only the center pixel, has a value S of 255 and each of its other neighbors, top left, top, top right and so forth, have values of 0. This sing~ r white pixel case col,.*)ollds to a small object. The result of the matrix multiplication and addition using the coefficient matrix is a value of 1/9 (255) or 28 for the center pixel, a value which is below the nomin~l threshold of 128. Now con~;~ler another case in which all the pixels have a value of 255 corresponding to a large object. Perforrning the low pass filtering operation on a 3 x 3 10 matrix for this case yields a value of 255 for the center pixel. Thus, large objects retain their values while small objects are reduced in amplitude or elimin~te~ In the plerelled method of operation, the low pass filtering process is p~,.ro..,lcd on the Ximage twice in succçc~ion.
In order to s~)aldLe; objects of interest, such as a tumor cell in the x image from other objects and background, a thresholding operation is performed designed to set pixels within cells of interest to a value of 255, and all other areas to 0. Thresholding ideally yields an image in which cells of interest are white and the r~m~inder of the image is black. A problem one faces in thresholding is where to set the threshold level. One cannot simply assume that cells of interest are indicated by any pixel value above the nominal threshold of 128. A typical im~ging system may use an inc~n~iescent halogen light bulb as a light source. As the bulb ages, the relative amounts of red and blue output can change. The tendency as the bulb ages is for the blue to drop off more than the red and the green. To accommodate for this light source variation over time, a dynamic thresholding process is used whereby the threshold is adjusted dyn~mically for each acquired image. Thus, for each 640 x 4g0 image, a single threshold value is derived specific to that image.
As shown in Fig. 18, the basic method is to calculate, for each field, the mean X value, and the standard deviation about this mean at 312. The threshold is then set at 314 to the mean plus an amount deflned by the product of a (user specified) factor and the standard deviation of the color converted pixel values. The standard deviation correlates to the structure and number of objects in the image. Preferably, the user specified factor is in the range of a~ in.al~ly 1.5 to 2.5. The factor is selected to be in the lower end of the range for slides in which the stain has primarily l~ ;,-e<~ within cell boundaries and the factor is selected to be in the upper end of the range for slides in which the stain is pervasively present throughout the slide. In this way, as areas are enco~ln~l~ d on the slide with greater or lower background int~n~iticc, the threshold may be raised or lowered to help reduce background objects. With this metllc d, the threshold changes in step with the W O 9712Q198 PC~nJS96/19166 aging of the light source such that the effects of the aging are cancelled out. The image matrix rPsllltin~ at 316 from the thresholding step is a binary image of black (O) and white (255) pixels.
As is often the case with thresholding operations such as that described above, some undesired areas will lie above the threshold value due to noise, small stained cell fragments, and 5 other artifacts. It is desired and possible to elimin~te these artifacts by virtue of their small size col.lpaled with legitimate cells of interest. Morphological processes are utilized to p~"rollil this function.
Morphological processing is similar to the low pass filter convolution process described earlier except that it is applied to a binary image. Similar to spatial convolution, the morphological 10 process L.d~lses an input image matrix, pixel by pixel, and places the processed pixels in an output matrix. Rather than calc~ ting a weighted sum of neighboring pixels as in the low pass convolution process, the morphological process uses set theory operations to combine neighboring pixels in a nonlinear fashion.
Erosion is a process ~I~ y a single pixel layer is taken away from the edge of an object.
15 Dilation is the opposite process which adds a single pixel layer to the edges of an object. The power of morphological proce~ing is that it provides for further discrimin~ti- n to elimin~te small objects that have survived the thresholding process and yet are not likely tumor cells. The erosion and dilation processes that make up a morphological "open" preferably make small objects disappear yet allows large objects to remain. Morphological ploces~h~g of binary images is described in detail in "Digital Image ~l~,ces~i.. g", pages 127-137, G.A. Baxes, John Wiley & Sons, (1994).
Fig. 19 illustrates the flow diagram for this process. As shown here, a morphological "open"
process p~ lls this suppression. A single morphological open consists of a single morphological erosion 320 followed by a single morphological dilation 322. Multiple "opens" consist of multiple erosions followed by multiple dilations. In the pler~ d embodiment, one or two morphological 25 opens are found to be suitable.
At this point in the processing chain, the processed image contains thresholded objects of interest, such as tumor ceils (if any were present in the original image), and possibly some residual artifacts that were too large to be elimin~hd by the processes above.
~ ig. 20 provides a flow diagram illustrating a blob analysis p~ .ned to ~l~qt~rmine the 30 number, size, and location of objects in the thresholded image. A blob is defined as a region of connected pixels having the same "color", in this case, a value of 255. Processing is performed over - the entire image to det.srmine the number of such regions at 324 and to del.,l.--ille the area and x,y coordinates for each detected blob at 326.

W O 97/20198 PCT~US96~19166 Co",pal;son of the size of each blob to a known minimlltn area at 328 for a tumor cell allows a refin~nent in decisions about which objects are objects of interest, such as tumor cells, and which are artifacts. The location (x,y coo,dinates) of objects identified as cells of interest in Lhis stage are saved for the final 40x reim~gin~ step described below. Objects not passing the size test are '5 di~ rded as artifacts.
The processing chain described above identifies objects at the sc~nning m~gnific~tion as cells of interest c~n~ t~ As illustrated in Fig. 21, at the completion of sç~nning the system hes to the 40x m~gnifi-~tion objective at 330, and each 5~n~ te is reimaged to confirm the identification 332. Each 40x image is leprucessed at 334 using the same steps as described above 10 but with test parameters suitably modified for the higher magnification (e.g. area). At 336, a region of interest centered on each confirmed cell is saved to the hard drive for review by the pathologist.
As noted earlier, a mosaic of saved images is made available for viewing by the pathologist.
As shown in Fig. 22, a series of images of cells which have been confirmed by the image analysis is ~ ese~lled in the mosaic 150. The pathologist can then visually inspect the images to make a d~ tion whether to accept (152) or reject (153) each cell image. Such a determination can be noted and saved with the mosaic of images for gcll.,ldlillg a printed report.
In addition to saving the image of the cell and its region, the cell coordindles are saved should the pathologist wish to directly view the cell through the oculars or on the image monitor. In this case, the pathologist reloads the slide carrier, selects the slide and cell for review from a mosaic 20 of cell images, and the system autom~fi~ ly positions the cell under the mi-luscope for viewing.
It has been found that normal cells whose nuclei have been stained with hematoxylin are often quite numerous, numbering in the thousands per 10x image. Since these cells are so numerous, and since they tend to clump, counting each individual nucleated cell would add an excessive processing burden, at the expense of speed, and would not l-~CP~ ily provide an ac~ dle count due 25 to clumping. The al)pal;us performs an estimation process in which the total area of each field that is stained hc.,ldluxylin blue is measured and this area is divided by the average size of a mlcle~ted cell. Fig. 23 outlines this process.
In this process, a single color band (the red channel provides the best co.lLldsl for blue stained nucleated cells) is plucessed by c~lcul~ting the average pixel value for each field at 342, establishing two threshold values (high and low) as indicated at 344, 346, and counting the number of pixels between these two values at 348. In the absence of dirt, or other opaque debris, this provides a count of the number of predominantly blue pixels. By dividing this value by the average area for a nucleated cell at 350, and looping over all fields at 352, an approximate cell count is o~tained. Preliminary testing of this process in-lic~tes an accuracy with +/- 15%. It should be noted W O 97~20198 PCT~U596/19166that for some slide ~ alalion techniques, the size of nucleated cells can be signific~ntly larger than the typieal size. The operator can select the appropriate nucleated eell size to con~ atd for these chal a.;~ i~lics .
As with any im~ging system, there is some loss of modulation transfer (i.e. contrast) due to S the mod~ tion transfer function (MTF) c~-~r~ct~ lics of the im~ing optics, eamera, electronies, and other colllponcnts. Since it is desired to save "high quality" images of cells of interest both for pathologist review and for arehival purposes, it is desired to co~ for these MTF losses.
An MTF colllpensdlion, or MTFC, is perforrned as a digital p~ocess applied to the acquired digital images. A digital filter is utilized to restore the high spatial frequency content of the images 10 upon storage, while m~int~ining low noise levels. With this MTFC technology, image quality is enh~nce~l or restored, through the use of digital processing methods as opposed to conventional oil-;...~..e..iion or other hardware based methods. MTFC is described further in "The Image Proe.o~sing Handbook," pages 225 and 337, J. (~. Rues, CRC Press (1995).
Referring to Fig. 24, the functions available in a user interface of the apparatus 10 are shown.
15 From the user interface, which is ~ svLcd graphically on computer monitor 26, an operator can select among a~dlalus functions which include acquisition 402, analysts 404, and system configuration 406. At the acquisition level 402, the ~cldL~r can select between manual 408 and automatic 410 modes of operation. In the manual mode, the operator is ~l~sellltid with manual operations 409. Patient h-ro.lllàlion 414 leg~l.li,lg an assay can be entered at 412.
In the analysis level 404, review 416 and report 418 functions are made available. At the review level 416, the Op~.ator can select a montage function 420. At this montage level, a pathologist can p~,rO...l ~ gnostic review functions including visiting an image 422, acceptlreject of cells 424, nucleated cell counting 426, acceptlreject of cell counts 428, and saving of pages at 430.
The report level 418 allows an Ope;ldlUI to g~ .al~; patient reports 432.
In the configuration level 406, the OpeldtOr can select to configure p.efere.. ces at 434, input Op~,ldl~ l infor nation 437 at 436, create a system log at 438, and toggle a menu panel at 440. The configuration pLcrerellces include scan area selection functions at 442, 452; m-~nt~e specifications at 444, bar code h~n-11ing at 446, default cell counting at 448, stain selection at 450, and scan objective selection at 454.

30 I;,~ .r.l~
While this invention has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form , W O 97/20198 PCTAUS96/19166and details may be made therein without departing firom the spirit and scope of the invention as defined by the appended claims.
What is claimed is:

Claims

1. A method of automatic analysis of a color image in a first color space of a magnified area of a slide having a biological specimen prepared with a great reagent characterized by the steps of:
transforming pixels of said image in said first color space to a second color space to differentiate candidate object of interest pixels from background pixels; and identifying candidate objects of interest from the candidate object of interest pixels in the second color space.

2. The method of Claim 1 wherein the first color space includes red, green, and blue components for each pixel and the transforming step includes forming a ratio between two components of the red, blue and green components for each pixel in the first color space to transform the pixels to the second color space.

3. The method of Claim 2 further comprising the step of:
selecting a grayscale value for each pixel in the second color space which corresponds to the ratio of components in the first color space.

4. The method of Claim 1 wherein the first color space includes red, green, and blue components for each pixel and the transforming step includes converting components of the red, blue and green components for each pixel in the first color space to pixel values in a hue, saturation, and intensity space.

5. The method of Claim 1 wherein the first color space includes red, green, and blue components for each pixel and the transforming step includes comparing pixel values for a single component for each pixel to a threshold to identify pixels having a component value equal to or greater than said threshold as candidate object of interest pixels and pixels having a component value less than the threshold as background pixels.

6. The method of claim 1 further comprising the steps of:
morphologically processing the candidate object of interest pixels to identify artifact pixels; and identifying the candidate objects of interest from the remaining candidate object of interest pixels not identified as artifact pixels.

7. The method of claim 6 further comprising the steps of:
filtering said candidate object of interest pixels with a low pass filter prior to morphologically processing the low pass filtered candidate object of interest pixels.

8. The method of claim 7 further comprising the steps of:
comparing said low passed filtered candidate object of interest pixels to a threshold prior to morphologically processing the candidate object of interest pixels which have values greater than or equal to the threshold value.

9. The method of claim 8 further comprising the steps of:
computing a mean value of said candidate object of interest pixels;

specifying a threshold factor;
computing a standard deviation for the candidate object of interest pixels; and setting the threshold to the sum of the mean value and the product of the threshold factor and the standard deviation prior to comparing the candidate object of interest pixels to the threshold.

10. The method of claim 7 the identifying further comprising the steps of:
grouping said morphologically processed candidate object of interest pixels into regions of connected candidate object of interest pixels to identify objects of interest;
comparing said objects of interest to blob analysis parameters; and storing location coordinates of the candidate objects of interest having an area corresponding to the blob analysis parameters.

11. The method of claim 10 wherein said previously performed method steps are performed on images acquired at low magnification and the method further comprising the steps of:
adjusting an optical system viewing the slide from which the objects of interest were identified to high magnification;
acquiring a high magnification image of the slide at the corresponding location coordinates each candidate object of interest;
transforming pixels of the high magnification image in the first color space to a second color space to differentiate high magnification candidate objects of interest pixels from background pixels; and identifying high magnification objects of interest from the candidate object of interest pixels in the second color space.

12. The method of claim 11 further comprising the steps of:
morphologically processing the high magnification candidate object of interest pixels to identify artifact pixels; and identifying the high magnification objects of interest from the remaining high magnification candidate object of interest pixels not identified as artifact pixels.

13. The method of claim 12 further comprising the steps of:
filtering said high magnification candidate object of interest pixels with a low pass filter prior to morphologically processing the low pass filtered high magnification candidate object of interest pixels.

14. The method of claim 13 further comprising the steps of:
comparing said low passed filtered high magnification candidate object of interest pixels to a threshold prior to morphologically processing the high magnification candidate object of interest pixels which have values greater than or equal to the threshold value.

15. The method of claim 14 further comprising the steps of:
computing a mean value of said high magnification candidate object of interest pixels;
specifying a threshold factor;
computing a standard deviation for the high magnification candidate object of interest pixels, and setting the threshold to the sum of the mean value and the product of the threshold factor and the standard deviation prior to comparing the high magnification candidate object of interest pixels to the threshold.

16. The method of claim 15 further comprising the steps of:
grouping said low passed filtered high magnification candidate object of interest pixels into regions of connected high magnification candidate object of interest pixels to identify high magnification objects of interest;

comparing said high magnification objects of interest to blob analysis parameters; and storing location coordinates of the high magnification objects of interest corresponding to the blob analysis parameters.

17. The method of claim 11 wherein an optical system is initially focused prior to performing the low magnification steps.

18. The method of claim 17 wherein the initial focusing of the optical system prior to performing the low magnification steps, further comprises the steps of:
(a) positioning the optical system at an initial Z stage position;
(b) acquiring at low magnification an image of a slide having a stained biological specimen thereon and calculating a pixel variance about a pixel mean for the acquired image;
(c) incrementing the position of the Z stage;
(d) repeating steps (b) and (c) for a fixed number of coarse iterations to form a first set of variance data;
(e) performing a least squares fit of the first set of variance data to a first function;
(f) positioning the Z stage at a position near the peak of the first function;
(g) repeating steps (b) and (c) for a fixed number of fine iterations to form a second set of variance data;
(h) performing a least squares fit of the second set of variance data to a second function;
(i) selecting the peak value of the least squares fit curve as an estimate of the best focal position; and (j) performing the above steps for an array of X-Y stage positions to form an array of focal positions and performing a least squares fit of the array of focal positions to yield a least squares fit focal plane.

19. The method of claim 17 wherein the initial focusing of the optical system prior to performing the low magnification steps, further comprises the steps of:
(a) positioning the optical system at an initial Z stage position;
(b) acquiring an image and calculating a pixel variance about a pixel mean for the acquired image;
(c) incrementing the position of the Z stage;
(d) repeating steps (b) and (c) for a fixed number of iterations;
(e) performing a least squares fit of the variance data to a known function; and (f) selecting the peak value of the least squares fit curve as an estimate of the best focal position.

20. The method of claim 11 the adjusting optical system step further comprising the steps of:
(a) positioning the optical system at an initial Z stage position;
(b) acquiring an image and selecting a center pixel of a candidate object of interest;
(c) defining a region of interest centered about the selected center pixel;
(d) preforming a fast fourier transform of said region of interest to identify frequency components for the region of interest and complex magnitudes for the frequency components;
(e) computing a power value by summing the square of the complex magnitudes for the frequency components that are within the range of frequencies of 25% to 75% of a maximum frequency component for the fast fourier transform of the region of interest;
(f) incrementing the position of the Z stage;
(g) repeating steps (b) - (e) for a fixed number of iterations; and (h) selecting the Z stage position corresponding to the largest power value as the best focal position.

21. The method of claim 11 the adjusting optical system step further comprising the steps of:

(a) positioning the optical system at an initial Z stage position;
(b) acquiring an image and selecting a center pixel of a candidate object of interest;
(c) defining a region of interest centered about the selected center pixel;
(d) applying a Hanning window function to the region of interest;
(d) performing a fast fourier transform of said region of interest following the application of the Hanning window function to identify frequency components for the region of interest and complex magnitudes for the frequency components;
(e) computing a power value by summing the square of the complex magnitudes for the frequency components for the fast fourier transform of the region of interest;
(f) incrementing the position of the Z stage;
(g) repeating steps (b) - (e) for a fixed number of iterations; and (h) selecting the Z stage position corresponding to the largest power value as the best focal position.

22. A method for initially focusing an optical system of an automated microscopy system Comprising the steps of:
(a) positioning an optical system at an initial Z stage position;
(b) acquiring an image and calculating a pixel variance about a pixel mean for the acquired image;
(c) incrementing the position of the Z stage;
(d) repeating steps (b) and (c) for a fixed number of iterations;
(e) performing a least squares fit of the variance data to a known function; and (f) selecting the peak value of the least squares fit curve as an estimate of the best focal position.

23. A method for initially focusing an optical system of an automated microscopy system comprising the steps of:
(a) positioning an optical system at an initial Z stage position;
(b) acquiring at low magnification an image of a slide having a stained biological specimen thereon and calculating a pixel variance about a pixel mean for the acquired image;
(c) incrementing the position of the Z stage;
(d) repeating steps (b) and (c) for a fixed number of coarse iterations to form a first set of variance data;
(e) performing a least squares fit of the first set of variance data to a first function;
(f) positioning the Z stage at a position near the peak of the first function;
(g) repeating steps (b) and (c) for a fixed number of fine iterations to form a second set of variance data;
(h) performing a least squares fit of the second set of variance data to a second function;
(i) selecting the peak value of the least squares fit curve as an estimate of the best focal position; and (j) performing the above steps for an array of X-Y stage positions to form an array of focal positions and performing a least squares fit of the array of focal positions to yield a least squares fit focal plane.

24. A method for focusing an optical system of an automated microscopy system at high magnification comprising the steps of:
(a) positioning an optical system at an initial Z stage position;
(b) acquiring an image and selecting a center pixel of a candidate object of interest;
(c) defining a region of interest centered about the selected center pixel;

(d) performing a fast fourier transform of said region of interest to identify frequency components for the region of interest and complex magnitudes for the frequency components;
(e) computing a power value by summing the square of the complex magnitudes for the frequency components that are within the range of frequencies of 25% to 75% of a maximum frequency component for the fast fourier transform of the region of interest;
(f) incrementing the position of the Z stage;
(g) repeating steps (b) - (e) for a fixed number of iterations; and (h) selecting the Z stage position corresponding to the largest power value as the best focal position.

25. A method for focusing an optical system of an automated microscopy system at high magnification comprising the steps of:
(a) positioning an optical system at an initial Z stage position;
(b) acquiring an image and selecting a center pixel of a candidate object of interest;
(c) defining a region of interest centered about the selected center pixel;
(d) applying a Hanning window function to the region of interest;
(d) performing a fast fourier transform of said region of interest following the application of the Hanning window function to identify frequency components for the region of interest and complex magnitudes for the frequency components;
(e) computing a power value by summing the square of the complex magnitudes for the frequency components for the fast fourier transform of the region of interest, (f) incrementing the position of the Z stage;
(g) repeating steps (b) - (e) for a fixed number of iterations; and (h) selecting the Z stage position corresponding to the largest power value as the best focal position.

26. Apparatus for automatic image analysis of a slide having a biological specimen, comprising:
an optical system having an X-Y stage;
means for scanning over a scan area of the slide at a plurality of locations at low magnification of the optical system;
means for acquiring a low magnification image at each location in the scan area;a processor for processing each low magnification image to detect candidate objects of interest;
means for storing X-Y coordinates of each location for each candidate object of interest;
means for adjusting the optical system to a high magnification;
means for repositioning the X-Y stage to the location for each candidate object of interest;
means for acquiring a high magnification image of each candidate object of interest; and a storage device for storing each high magnification image.

27. Apparatus for automatic image analysis of a slide having a biological specimen, comprising:
an optical system having an X-Y stage;
means for scanning over a scan area of the slide at a plurality of locations at low magnification of the optical system;
means for acquiring a low magnification image at each location in the scan area;a processor for processing each low magnification image to detect candidate of objects of interest;
means for storing X-Y coordinates of each location for each candidate of object of interest;
means for adjusting the optical system to a high magnification;

means for repositioning the X-Y stage to the location for each candidate object of interest;
means for acquiring a high magnification image of each candidate of object of interest; and a storage device for storing each high magnification image.