 |
Claims  |
|
|
What is claimed is:
1. An automated cytological specimen classifier for classifying cells
contained in a smear on a slide to identify cells that are likely to be
malignant or pre-malignant, comprising:
(a) microscope means for obtaining a vie of at least part of a cytological
specimen including cells and other material located generally randomly on
a slide in an arrangement which can include other than a single layer of
cells,
(b) camera means for creating an image of such view,
(c) image digitizing means for producing a digital representation of such
image,
(d) a primary classifier means for detecting objects in a digital
representation of a cytological specimen based on a detectable feature,
said primary classifier means comprising a classifier means for detecting
cells that are likely to be malignant or pre-malignant as well as other
cells and material that initially appear to have characteristics of a
malignant cell or a pre-malignant cell based on integrated optical
density, and
(e) a secondary classifier for distinguishing pre-malignant and malignant
cells form other cells and material among the objects detected by the
primary classifier, said secondary classifier means comprising a neural
computer apparatus means for effecting such distinguishing as a function
of training thereof.
2. The automated classifier of claim 1 wherein said primary classifier
means comprises a statistical classifier.
3. The automated classifier of claim 1 wherein such cytological specimen
includes overlapping cells.
4. The automated classifier of claim 1 wherein said secondary classifier
means is operable to distinguish pre-malignant and malignant cells among
overlapping arrangements of cells and other material.
5. The automated classifier of claim 1, wherein said camera means is
positioned to create an image of a portion of such cytological specimen
from such view.
6. The automated classifier of claim 1, said neural computer apparatus
means comprising an electronic neural computer.
7. The automated classifier of claim 1 wherein the primary classifier means
is restricted to evaluation of the cellular nucleus while the secondary
classifier means evaluates both the cellular nucleus and its surrounding
cytoplasm.
8. The automated classifier of claim 1 wherein both the primary classifier
means and the secondary classifier means are restricted to evaluation of
the cellular nucleus.
9. The automated classifier of claim 1 further comprising means for making
an additional non-neural net classification of nuclear morphological
components in addition to integrated optical density, said means being
coupled between the primary classifier means and the secondary classifier
means.
10. The automated classifier of claim 1, said microscope means comprising
an automated microscope.
11. The automated classifier of claim 1, said camera means comprising a
video camera.
12. The automated classifier of claim 1, said camera means comprising a
charge coupled device.
13. The automated classifier of claim 1, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen which has a feature that exceeds a threshold
level.
14. The automated classifier of claim 1, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen which has a feature that exceeds a threshold
integrated optical density.
15. The automated classifier of claim 1, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen based on morphological criteria.
16. The automated classifier of claim 1 further comprising a neural network
pre-screening classifier means for recognition and classification of
general areas within the digital representation of a specimen that contain
material other than a cellular monolayer prior to primary classification
by said primary classifier means.
17. The automated classifier of claim 16 wherein the output form the
pre-screening classifier means is utilized to exclude such areas from
further analysis.
18. The automated classifier of claim 16 wherein the output form the
pre-screening classifier means is utilized to modify further analysis of
images found within such areas.
19. A method of classifying cytological specimens, comprising using a
primary classifier apparatus primarily classifying a specimen which is
generally randomly arranged and can include other than in a single layer
to determine locations of interest, and secondarily classifying such
locations of interest using a neural network computer apparatus.
20. The method of claim 19, wherein said primarily classifying step
comprises using a video camera or charge coupled device (CCD) to obtain
images of the specimen, a digitizer to digitize such images and an
integrated optical density detector.
21. The method of claim 19, wherein said primary classifying comprises
using a neural network computer apparatus.
22. The method of claim 19, wherein said step of using a primary classifier
apparatus primarily classifying a specimen comprises using a statistical
classifier.
23. The method of claim 19, wherein said step of using a primary classifier
apparatus primarily classifying a specimen comprises making a
classification based on morphology.
24. The method of claim 19, wherein said step of using a primary classifier
apparatus primarily classifying a specimen comprises making such primary
classification based on integrated optical density.
25. The method of claim 19, further comprising training such neural network
computer apparatus to identify cytological specimens of interest.
26. An automated cytological specimen classifier, comprising:
(a) microscope means for obtaining a view of at least part of a cytological
specimen including cells and other material located generally randomly in
an arrangement which can include other than a single layer of cells,
(b) camera means for creating an image of such view,
(c) image digitizing means for producing a digital representation of such
image,
(d) primary classifier means for detecting objects in a digital
representation of a cytological specimen based on a detectable feature,
said primary classifier means comprising a classifier for detecting cells
that are likely to be of a predetermined cell type as well as other cells
and material that initially appear to have characteristics of such
predetermined cell type, and
(e) secondary classifier means for distinguishing cells of such
predetermined cell type from other cells and material among the objects
detected by said primary classifier means, said secondary classifier means
comprising a neural computer apparatus means for effecting such
distinguishing as a function of training thereof.
27. The automated classifier of claim 16, wherein the primary classifier
means is restricted to evaluation of the cellular nucleus while the
secondary classifier means evaluates both the cellular nucleus and its
surrounding cytoplasm.
28. The automated classifier of claim 26 wherein both the primary
classifier means and the secondary classifier means are restricted to
evaluation of the cellular nucleus.
29. The automated classifier of claim 26, further comprising means for
making an additional non-neural net classification of nuclear
morphological components, said means being coupled between the primary
classifier means and said secondary classifier means.
30. The automated classifier of claim 26, said microscope means comprising
an automated microscope.
31. The automated classifier of claim 16, said camera means comprising a
video camera.
32. The automated classifier of claim 26, said camera means comprising a
charge coupled device.
33. The automated classifier of claim 26, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen which have a feature that exceeds a threshold
level.
34. The automated classifier of claim 26, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen which has a feature that exceeds a threshold
integrated optical density.
35. The automated classifier of claim 26, said primary classifier means
comprising means for detection of objects in such digital representation
of a cytological specimen based on morphological criteria.
36. The automated classifier of claim 26, wherein said camera means is
positioned to create an image of a portion of such cytological specimen
from such view.
37. The automated classifier of claim 26, said neural computer apparatus
means comprising an electronic neural computer.
38. The automated classifier of claim 16 wherein primary classifier means
comprises a statistical classifier.
39. The automated classifier of claim 16 wherein such cytological specimen
includes overlapping cells.
40. The automated classifier of claim 16 wherein said secondary classifier
means is operable to distinguish cells of such predetermined cell type
among overlapping arrangements of cells and other materials.
41. The automated classifier of claim 26 further comprising a neural
network pre-screening classifier means for identifying general areas
within the digital representation of a specimen that contain material
other than a cellular monolayer prior to primary classification.
42. The automated classifier of claim 41 wherein the output from the
pre-screening classifier is utilized to exclude such identified areas from
further analysis.
43. The automated classifier of claim 41 wherein the output from the
pre-screening classifier means is utilized to modify further analysis of
images found within the areas of the specimen identified by said
pre-screening classifier means. |
|
 |
|
 |
Claims |
|
|
|
Description |
|
|
TECHNICAL FIELD
This invention relates generally, as indicated, to cell classification and,
more particularly, to the use of neural networks and/or neurocomputers for
increasing the speed and accuracy of cell classification.
BACKGROUND OF THE INVENTION
The cervical smear (Pap test) is the only mass screening cytological
examination which requires visual inspection of virtually every cell on
the slide. The test suffers from a high false negative rate due to the
tedium and fatigue associated with its current manual mode of performance.
Cell classification is typically performed on a "piece-work" basis by
"cytotechnicians" employed by pathology laboratories and in some
circumstances by salaried technicians. Due to the clearly life threatening
nature of the false negative problem with its resultant possibility of
undiagnosed cervical cancer, the American Cancer Society is considering
doubling the frequency of recommended Pap smears. This, however, will
certainly overload an already overburdened cervical screening industry as
increasingly fewer individuals are willing to enter the tedious and
stressful field of manual cervical smear classification. An American
Cancer Society recommendation to increase Pap smear frequency may only
serve to increase the false negative rate by decreasing the amount of time
spent on manual examination of each slide. A thorough manual examination
should take no less than fifteen minutes per slide although a
cytotechnician, especially one under a heavy workload, may spend less than
half this amount of time. The College of American Pathology is well aware
of this problem and would rapidly embrace an automated solution to
cervical smear screening.
Due to the clear commercial potential for automated cervical smear analysis
several attempts to this end have been made in the prior art. These
attempts have proven to be unsuccessful since they have relied exclusively
on classical pattern recognition technology (geometric, syntactic,
template, statistical) or artificial intelligence (AI) based pattern
recognition, i.e., rule-based expert systems. There is, however, no clear
algorithm or complete and explicit set of rules by which the human
cytotechnician or pathologist uses his experience to combine a multitude
of features to make a classification in gestalt manner. Cervical smear
classification is, therefore, an excellent application for neural network
based pattern recognition.
An example of the limitations of the prior art can be found in the 1987
reference entitled "Automated Cervical Screen Classification" by Tien et
al, identified further below.
Background references of interest are, as follows:
Rumelhart, David E. and McClelland, James L., "Parallel Distributed
Processing," MIT Press, 1986, Volume 1;
Tien, D. et al, "Automated Cervical Smear Classification," Proceedings of
the IEEE/Ninth Annual Conference of the Engineering in Medicine and
Biology Society, 1987, p. 1457-1458;
Hecht-Nielsen, Robert, "Neurocomputing: Picking the Human Brain," IEEE
Spectrum, March, 1988, p. 36-41; and
Lippmann, Richard P., "An Introduction to Computing with Neural Nets," IEEE
ASSP Magazine, April, 1987, p. 4-22.
BRIEF SUMMARY OF THE INVENTION
It is, therefore, a principal object of the present invention to provide an
automated system and method for the classification of cytological
specimens into categories, for example, categories of diagnostic
significance.
Briefly, the invention includes an initial classifier (sometimes referred
to as a primary classifier) preliminarily to classify a cytological
specimen and a subsequent classifier (sometimes referred to as a secondary
classifier) to classify those portions of the cytological specimen
selected by the initial classifier for subsequent classification, wherein
the subsequent classifier includes a neural computer or neural network.
In one embodiment the primary classifier may include a commercially
available automated microscope in the form of a standard cytology
microscope with a video camera or CCD array with the microscope stage
controlled for automatic scanning of a slide. Images from the camera are
digitized and outputted to the secondary classifier in the form of a
computer system. The computer system includes a neural network as defined
below and is disclosed, too, in several of the references referred to
herein, which is utilized in the performance of cell image identification
and classification into groups of diagnostic interest. In an alternate
embodiment the primary classifier may include a neural network. Other
alternate embodiments also are disclosed below.
It is a further object of the present invention that it perform its
classification of a group of specimens within the period of time typically
consumed for this task by careful manual screening (i.e., approximately 15
minutes/specimen).
It is a further object of the present invention that it perform its
classification on cytological specimens which contain the numbers and
types of objects other than single layers of cells of interest that are
typically found in cervical smears (e.g., clumps of cells, overlapping
cells, debris, leucocytes, bacteria, mucus).
It is a further object of the present invention to perform the
above-described classification on cervical smears for the detection of
pre-malignant and malignant cells.
It is a further object of the present invention that it perform its
classification with smaller false negative error rates than typically
found in conventional manual cervical smear screening.
An advantage of the cytological classification system of the present
invention is that classification of cytological specimens into medically
significant diagnostic categories will be more reliable, i.e., with lower
false negative error rates.
A further advantage of the cytological classification system of the present
invention is that it does not require a modification in the procedure by
which cellular specimens are obtained from the patient.
A further advantage of the cytological classification system of the present
invention is that it will permit reliable classification within processing
time constraints that permit economically viable operation.
These and other objects, advantages and features of the present invention
will become evident to those of ordinary skill in the art after having
read the following detailed description of the preferred embodiment.
It is noted here that the published articles cited herein are specifically
incorporated by reference.
Moreover, it is noted here that the invention is described herein mainly
with respect to classification of cytological specimens in the form of a
cervical smear, e.g., as typically is done in connection with a Pap test.
However, it will be appreciated that this is but one example of the
application of the principles of the invention which are intended for
application for classification of many other cytological specimens.
BRIEF DESCRIPTION OF THE DRAWINGS
In the annexed drawings:
FIG. 1 is a block diagram for a neural network based automated cytological
specimen screening device in accordance with the present invention;
FIG. 2 is a representation of a three-layer neural network of the type
utilized in the preferred embodiment;
FIG. 3 is a block diagram of the alternate embodiment of the automated
screening device in accordance with the present invention;
FIG. 4 is a block diagram of an alternate embodiment of the automated
screening device in accordance with the present invention;
FIG. 5 is a block diagram of an alternate embodiment of the automated
screening device in accordance with the present invention;
FIG. 6 is a block diagram of an alternate embodiment of the automated
screening device in accordance with the present invention; and
FIG. 7 is a block diagram of an alternate embodiment of the automated
screening device in accordance with the present invention.
DESCRIPTION OF THE PREFERRED AND ALTERNATE EMBODIMENTS
FIG. 1 illustrates a neural network based automated cytological specimen
screening device in accordance with the present invention and referred to
by the general reference numeral 10. The classification device 10 includes
an automated microscope 11, a video camera or CCD device 12, an image
digitizer 13, and classifier stages 14, and 15.
The automated microscope 11 effects relative movement of the microscope
objective and the specimen, and video camera or CCD 12 obtains an image or
picture of a specific portion of the cytological specimen. The image is
digitized by the image digitizer 13 and the information therefrom is
coupled to the classifier 14. In the preferred embodiment, classifier 14
is a commercially available statistical classifier which identifies cell
nuclei of interest by measurement of their integrated optical density
(nuclear stain density). This is the sum of the pixel grey values for the
object, corrected for optical errors. Compared to normal cells, malignant
cells tend to possess a larger, more densely staining nucleus.
Objects which pass classifier 14 consist of pre-malignant and malignant
cells but also include other objects with high integrated optical density
such as cell clumps, debris, leucocytes and mucus. The task of the
secondary classifier 15 is to distinguish pre-malignant and malignant
cells from these other objects.
A neural network is utilized to implement secondary classifier 15. Detailed
descriptions of the design and operation of neural networks suitable for
implementation of secondary classifier 15 can be found in the references
cited herein. A brief description of this information is provided below.
Based on the data obtained by the primary classifier for the cytological
specimen, the secondary classifier is used to check specific areas of the
specimen that are, for example, determined to require further screening or
classification. Such further examination by the secondary classifier may
be effected by reliance on the already obtained digitized image data for
the selected areas of the specimen or by the taking of additional data by
the components 11-13 or by other commercially available optical or other
equipment that would provide acceptable data for use and analysis by the
secondary classifier 15.
A neural network is a highly parallel distributed system with the topology
of a directed graph. The nodes in neural networks are usually referred to
as "processing elements" or "neurons" while the links are generally known
as "interconnects." Each processing element accepts multiple inputs and
generates a single output signal which branches into multiple copies that
are in turn distributed to the other processing elements as input signals.
Information is stored in the strength of the connections known as weights.
In an asynchronous fashion, each processing element computes the sum of
products of the weight of each input line multiplied by the signal level
(usually 0 or 1) on that input line. If the sum of products exceeds a
preset activation threshold, the output of the processing element is set
to 1, if less, it is set to 0. Learning is achieved through adjustment of
the values of the weights.
For the present invention, the preferred embodiment is achieved by
utilization of a three-layer neural network of the type described in the
Lippman reference as a "multi-layer perception" and discussed in detail in
Chapter 8 of the Rumelhart reference. Other types of neural network
systems also may be used.
A three-layer neural network consists of an input layer, an output layer,
and an intermediate hidden layer. The intermediate layer is required to
allow for internal representation of patterns within the network. As shown
by Minsky and Papert in their 196 book entitled "Perceptrons" (MIT Press),
simple two-layer associative networks are limited in the types of problems
they can solve. A two-layer network with only "input" and "output"
processing elements can only represent mappings in which similar input
patterns lead to similar output patterns. Whenever the real world problem
is not of this type, a three-layer network is required. It has been shown
that with a large enough hidden layer, a threelayer neural network can
always find a representation that will map any input pattern to any
desired output pattern. A generic three-layer neural network of the type
utilized in the preferred embodiment is shown in FIG. 2.
Several important features of neural network architectures distinguish them
from prior art approaches to the implementation of classifier 15.
1. There is little or no executive function. There are only very simple
units each performing its sum of products calculation. Each processing
element's task is thus limited to receiving the inputs from its neighbors
and, as a function of these inputs, computing an output value which it
sends to its neighbors. Each processing element performs this calculation
periodically, in parallel with, but not synchronized to, the activities of
any of its neighbors.
2. All knowledge is in the connections. Only very short term storage can
occur in the states of the processing elements. All long term storage is
represented by the values of the connection strengths or "weights" between
the processing elements. It is the rules that establish these weights and
modify them for learning that primarily distinguish one neural network
model from another. All knowledge is thus implicitly represented in the
strengths of the connection weights rather than explicitly represented in
the states of the processing elements.
3. In contrast to algorithmic computers and expert systems, the goal of
neural net learning is not the formulation of an algorithm or a set of
explicit rules. During learning, a neural network self-organizes to
establish the global set of weights which will result in its output for a
given input most closely corresponding to what it is told is the correct
output for that input. It is this adaptive acquisition of connection
strengths that allows a neural network to behave as if it knew the rules.
Conventional computers excell in applications where the knowledge can be
readily represented in an explicit algorithm or an explicit and complete
set of rules. Where this is not the case, conventional computers encounter
great difficulty. While conventional computers can execute an algorithm
much more rapidly than any human, they are challenged to match human
performance in non-algorithmic tasks such as pattern recognition, nearest
neighbor classification, and arriving at the optimum solution when faced
with multiple simultaneous constraints. If N exemplar patterns are to be
searched in order to classify an unknown input pattern, an algorithmic
system can accomplish this task in approximately order N time. In a neural
network, all of the candidate signatures are simultaneously represented by
the global set of connection weights of the entire system. A neural
network thus automatically arrives at the nearest neighbor to the
ambiguous input in order 1 time as opposed to order N time.
For the present invention, the preferred embodiment is achieved by
utilization of a three-layer backpropagation network as described in the
Rumelhart reference for the neural network of classifier stage 15.
Backpropagation is described in detail in the Rumelhart reference. Briefly
described, it operates as follows. During net training, errors (i.e., the
difference between the appropriate output for an exemplar input and the
current net output for that output) are propagated backwards from the
output layer to the middle layer and then to the input layer. These errors
are utilized at each layer by the training algorithm to readjust the
interconnection weights so that a future presentation of the exemplar
pattern will result in the appropriate output category. Following the net
training, during the feed-forward mode, unknown input patterns are
classified by the neural network into the exemplar category which most
closely resembles it.
The output of neural net classifier 15 indicates the presence or absence of
pre-malignant or malignant cells. The location of the cells on the input
slide is obtained from X-Y plane position coordinates outputted
continually by the automated microscope. This positional information is
outputted to printer or video display 17 along with diagnosis and patient
identification information so that the classification can be reviewed by a
pathologist.
In the preferred embodiment, the parallel structure of the neural network
is emulated by execution with pipelined serial processing as performed by
one of the commercially available neurocomputer accelerator boards. The
operation of these neurocomputers is discussed in the Spectrum reference
cited. The neural network preferably is a "Delta" processor, which is a
commercially available neurocomputer of Science Applications International
Corp. (SAIC) (see the Hecht-Nielsen reference above) that has demonstrated
a sustained processing rate of 10.sup.7 interconnects/second in the
feed-forward (i.e., non-training) mode. For a typical cervical smear
containing 100,000 cells, 1-2% of the cells or approximately 1,500 images
will require processing by classifier 15. As an example of the data rates
which result, assume that following data compression an image 50 .times.
50 pixels is processed by classifier 15. The input layer for the neural
network, therefore, consists of 2,500 processing elements or "neurons."
The middle layer consists of approximately 25% of the input layer, or 625
neurons. (The number of output neurons is equal to the number of
diagnostic categories of interest. This small number does not
significantly affect this calculation.) The number of interconnects is
thus (2500)(625) or approximately 1.5 .times. 10.sup.6. At a processing
rate of 10 interconnects/second, the processing by classifier 15 of the
1,500 images sent to it by classifier 14 will take less than four minutes.
Currently available embodiments of classifier 14 operate at a rate of
50,000 cells/minute (refer to the Tien et al citation). With classifier 14
operating at a rate of 50,000 cells/minute, the four minutes consumed by
classifier 15 is added to the two minutes used by classifier 14 for a
total of six minutes to analyze the 100,000, cell images on the slide. As
discussed above, an accurate manual cervical smear analysis takes
approximately 15 minutes/slide. Prior art automated attempts using a
non-neural network embodiment of classifier 15 require over one
hour/slide. This example is not meant in any way to limit the actual
configuration of the present invention, but rather to demonstrate that it
is capable of achieving the object of processing cervical smears and other
cytological samples within the time period required for commercially
feasible operation.
In the preferred embodiment, primary classifier 14 is restricted to
evaluation of the cellular nucleus while the secondary classifier 15
evaluates both the nucleus and its surrounding cytoplasm. The ratio
between the nucleus and cytoplasm is an important indicator for
pre-malignant and malignant cell classification. In an alternate
embodiment, both classifier 14 and classifier 15 are limited to evaluation
for the cellular nuclei.
Output information from the secondary classifier 15 is directed to an
output monitor and printer 17, which may indicate a variety of information
including, importantly, whether any cells appear to be malignant or
pre-malignant, appear to require further examination, etc.
FIG. 3 illustrates an alternate embodiment in which an additional neural
net classifier stage 16 is added to preprocess the slide for large areas
of artifactual material, i.e., material other than single layer cells of
interest. This includes clumps of cells, debris, mucus, leucocytes, etc.
Positional information obtained in this pre-screen is stored for use by
the remainder of the classification system. The information from
classifier stage 16 is utilized to limit the processing required by
classifier 15. Classifier stage 14 can ignore all material within the
areas defined by the positional coordinates outputted by classifier 16.
This will result in less information being sent for processing by
classifier 15. A diagnosis is, therefore, made on the basis of
classification of only those cells which lie outside of these areas. If an
insufficient sample of cells lies outside of these areas for a valid
diagnosis, this information will be outputted on 17 as an "insufficient
cell sample."
FIG. 4 illustrates an alternate embodiment in which the images within the
areas identified by classifier 16 are not ignored but are instead
processed by a separate classifier 18 which operates in parallel with
classifier 15. The training of the neural net which composes classifier 18
is dedicated to the distinction of pre-malignant and malignant cells from
said artifactual material.
FIG. 5 illustrates an alternate embodiment wherein an additional non-neural
net classification of nuclear morphological components, exclusive of
integrated optical density, is placed between classifier 14 and classifier
15. This classification is performed by classifier 19.
FIG. 6 illustrates an alternate embodiment in which a commercially
available SAIC neurocomputer is optimized for feed-forward processing 20.
Through deletion of learning-mode capacity, all neurocomputer functions
are dedicated to feedforward operation. Learning is completed on a
separate unmodified neurocomputer which contains both the learning and
feed-forward functions.
Following the completion of learning, the final interconnection weights are
transferred to the optimized feedforward neurocomputer 20. Dedication of
neurocomputer 20 to the feed-forward mode results in a sustained
feed-forward operation rate of 10.sup.8 interconnects/second vs. 10.sup.7
interconnects/second for the non-optimized board as commercially supplied.
The optimized feed-forward neural network 20 is utilized to perform the
functions of classifiers 14 and 16 in FIGS. 1, 3, 4, and 5. By utilizing
neural net classifier 20 to perform the function of statistical classifier
14, cells of interest which are not necessarily malignant cervical cells,
and which do not therefore exceed the integrated optical density threshold
of classifier 14, would nevertheless be detected. An example would be the
detection of endometrial cells which, while not necessarily indicative of
cervical malignancy, are indicative of uterine malignancy when found in
the Pap smear of a post-menopausal patient.
As an example of the data rates that result from this embodiment in FIG. 6,
assume outside slide dimensions of 15 mm .times. 45 mm or a total slide
area of 675 .times. 10.sup.6 micrometers.sup.2. Neural net 20 processes a
sliding window over this area for analysis. This window has dimensions of
20 micrometers .times. 20 micrometers or an area of 400 micrometers.sup.2.
There are, therefore, 1.5 .times. 10.sup.6 of these windows on the 15 mm
.times. 45 mm slide. For the primary classification function performed by
neural net 20, a resolution of 1 micrometer/pixel is sufficient to detect
those objects which must be sent to secondary neural network classifier 15
for further analysis. The input pattern for the image window analyzed by
classifier 20 is therefore 20 .times. 20 pixels or 400 neurons to the
input layer of neural net 20. The middle layer consists of approximately
25% of the input layer or 100 neurons. (As discussed above in the data
rate calculation for classifier 15, the number of output layer neurons is
small and does not significantly affect our results.) The number of
interconnections in classifier 20 is thus approximately (400)(100) or 40
.times. 10.sup.3. At a processing rate of 10.sup.8 interconnects/second,
each image from the sliding window will take 400 microseconds for neural
net 20 to classify. In a 15 mm .times. 45 mm slide, there are 1.5 .times.
10.sup.6 of the 400 micrometer.sup.2 windows which require classification
by neural net 20. Total classification time for neural net 20 is therefore
(1.5 .times. 10.sup.6)(400 .times. 10.sup.-6) = 600 seconds or ten
minutes. If this ten minutes is added to the approximately four minutes
required for secondary neural net classifier 15, a total of 14
minutes/slide results. This example is not meant in any way to limit the
actual configuration of the present invention, but rather to demonstrate
that it is capable of achieving the object of processing cervical smears
and other cytological samples within the time period required for
commercially feasible operation.
Speed of processing data can be enhanced, too, by using parallel
processing. For example, plural commercially available neurocomputers from
SAIC can be coupled to effect parallel processing of data, thus increasing
overall operational speed of the classifier using the same.
FIG. 7 illustrates an alternate embodiment in which neural net primary
classifier 20 is utilized in conjunction with, rather than as a substitute
for morphological classification and area classification. By dedication of
classifier 20 to the detection of those few cell types which are of
interest, but which cannot be detected by other means, the resolution
required of classifier 20 is minimized.
Although the present invention has been described in terms of the presently
preferred embodiment, it is to be understood that such disclosure is not
to be interpreted as limiting. Various alterations and modifications will
no doubt become apparent to those skilled in the art after having read the
above disclosure. Accordingly, it is intended that the appended claims be
interpreted as covering all alterations and modifications as fall within
the true spirit and scope of the invention.
* * * * *
|
|
|
|
|
|
|
Description |
|
