 |
Description  |
|
|
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention involves a system and a method for automatically measuring a
length or other distance parameter of a body structure based on an image
of the structure, in particular, where the structure is imaged using
ultrasound.
2. Background of the Invention
During ultrasonic examinations, clinicians often want to measure some
feature of the patient's body. This is particularly common in obstetric
examinations where the sonographer often wishes to measure such things as
the fetus's femur length (FL), humerus length (HL), head circumference
(HC), abdominal circumference (AC), occipitofrontal diameter (OFD--the
length of the line segment that lies between the left and right halves of
the brain and connects opposing points of the skull), and biparietal
diameter (BPD--the longest line segment with endpoints on the midpoints of
the skull that is perpendicular to the line of the OFD).
There are, accordingly, several known ultrasound-based devices that
incorporate some way to measure linear or arc length of structures in a
patient's body. In most of these known systems, the user first looks at
the ultrasound machine's display screen to determine which portion
corresponds to the structure of interest. She then moves a trackball or
mouse to position a cursor along this displayed structure and "clicks" on
or otherwise marks various points along the displayed image. The
processing system then "connects the dots" in software to form an
approximate representation of the structure. and estimates the length
according to some predetermined measure. Another common procedure is to
mark a diameter of an approximating ellipse and to then use a repeat
toggle to "open" the ellipse to approximate the circumference of a
structure.
One big disadvantage of such known systems is that it takes a lot of time
for the operator to define the structure of interest--in order to get a
usefully accurate representation of, say, the fetus's head, the user may
need to mark tens of points. Studies of obstetric sonography have
indicated, for example, that 20-30% of the operator's time is taken up by
performing routine measurements. Moreover, the accuracy of the
measurements will depend on how carefully the user marks the displayed
structure of interest and it is known that measurement results can vary
greatly depending on the sonographer.
One way that has been proposed to speed up the measurement process is to
automate it, allowing the ultrasound machine's processing system itself to
identify and then measure the structure of interest. Common to such
proposals, however, is that they treat obstetric ultrasound images as any
other images, and they apply conventional image-processing techniques to
extract image features for measurements. These approaches ignore the fact
that it takes a great deal of computational effort for a system to
identify structure that a human viewer can identify at a glance, often
much more accurately than the machine, especially in the presence of
significant image noise. Furthermore, the accuracy and robustness of these
systems is questionable since image features can change significantly from
one image to another, and can deteriorate rapidly when image quality is
poor.
These proposals for fully automatic identification and measurement thus
ignore how human operators can consistently perform these measurements,
even for images with poor quality. For example, abdominal circumference
(AC) is one of the most difficult obstetric measurements because of poor
tissue boundary definition, yet human operators can usually readily
identify the structure and mark reference points for the measurement
routines.
Yet another disadvantage of known systems is that they use approximating
functions such as best-fit circles, ellipses and line segments that
introduce more error than is desirable--few heads have a perfectly
circular or elliptical cross-section, and few femurs are perfectly
straight. Deviations from the assumed ideal translate to measurement
errors.
What is needed is a way to identify and measure body structures fast, but
that still incorporates the user's ability to quickly identify features
visually as well as other experiential knowledge of the shape of the
structures of interest.
SUMMARY OF THE INVENTION
According to the invention, human body structures, including those of a
fetus, are automatically measured using ultrasound by first using an
ultrasonic transducer or prestored ultrasound scan to generate an image
frame as a pattern of pixels, each pixel having a brightness value
corresponding to an echo signal from a corresponding portion of an
interrogation region of the patient's body, which includes the body
structure. The image frame is displayed on a screen and includes a
structure frame portion that corresponds to the body structure.
The user then designates a general geometry feature of the displayed body
structure and at least one measurement parameter associated with the
designated geometry feature. For curved, closed structures such as the
head or abdomen, the measurement parameters may, for example, be the HC,
AC, OFD, or BPD. For mainly straight structures such as the femur or
humerus, the measurement parameter will normally be the end-to-end length.
Next, the user selects at most two reference points associated with the
displayed body structure.
A processing system then filters the displayed image to identify the
structure frame portion, generates an approximating function corresponding
to the designated measurement parameter, and calculates each measurement
parameter as a predetermined function of the approximating function.
The calculated measurement parameters are then preferably displayed or
otherwise recorded so that the user can see and evaluate them.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a block diagram of an ultrasonic imaging system with an
identification and measuring processor and display according to the
invention.
FIGS. 2 and 3 illustrate an ultrasonic scan of a cross section of a head
without and with, respectively, user-selected reference points and
system-generated delimiting curves.
FIG. 4 illustrates raster-to-polar conversion of a delimited portion of the
image in FIGS. 2 and 3.
FIG. 5 shows the polar image of FIG. 4 in a binarized form.
FIG. 6 illustrates the operation of certain morphologic filtering rules
applied to binarized images.
FIGS. 7 and 8 illustrate an ultrasonic scan of a mainly straight body
structure such as a femur without and with, respectively, a user-selected
reference point and system-generated windows.
FIG. 9 illustrates a delimited and oriented (rotated) portion of the image
in FIGS. 7 and 8.
FIG. 10 shows the oriented image portion FIG. 9 in a binarized form.
DETAILED DESCRIPTION
FIG. 1 illustrates the main components of an ultrasonic imaging system
according to the invention. The user enters various conventional scan
parameters into an input unit 100, which typically includes such devices
as a keyboard, knobs, and buttons, and a cursor-control device such as a
trackball 101 or mouse. The input unit is connected to a processing system
102, which will typically be an electrically connected and cooperating
group of processors such as microprocessors, digital signal processors,
and application-specific integrated circuits (ASIC); the processing system
may, however, also be implemented by a single processor as long as it is
fast enough to handle the various tasks described below.
As in known systems, the processing system 102 sets, adjusts, and monitors
the operating parameters of a conventional transmission control circuit
104, which generates and applies electrical control and driving signals to
an ultrasonic probe 106, which includes an array of piezoelectric
elements. As is well known in the art, the piezoelectric elements generate
ultrasonic waves when electrical signals of the proper voltage and
frequency are applied to them.
By placing the probe 106 against the body of a patient, these ultrasonic
waves enter a portion 110 of the patient's body. By varying the phasing,
amplitude, and timing of the driving signals, the ultrasonic waves are
focussed to form a series of scan lines 112 that typically fan out from
the probe. Several such scan lines are shown extending into the patient's
body in FIG. 1. A region of interest, that is, the region that the user
wants to have an image of, is shown as an interrogation region or volume
114. The manner in which ultrasonic scanning signals are controlled,
generated, and applied to a patient's body is well understood in the art
and is therefore not described further.
Ultrasonic echoes from the waves transmitted into the body return to the
array in the probe 106. As is well understood, the piezoelectric elements
in the array thereby convert the small mechanical vibrations caused by the
echoes into corresponding electrical signals. Amplification and other
conventional signal conditioning is then applied to the return signals by
a reception controller 116. This processing includes, as needed, such
known signal conditioning as time-gating, gain compensation, and noise
filtering, in order to identify the echo signals that correspond to the
interrogation region 114.
The reception controller 116, all or part of which is normally integrated
into the processing system 102 itself, processes the ultrasonic,
radio-frequency (RF) echo signals from the array (typically on the order
of a few to tens of megahertz) to form reception beams along the
transmission beam direction. This is well known in the art of ultrasonic
imaging. The magnitude values of the received beams for the
two-dimensional interrogation region are stored digitally in a memory 118
as 2-D frame data 120. Each set of frame data corresponds to one image
frame, that is, to a 2-D cross section of the interrogation region.
The stored data format is normally not in the same shape or size as what
the user wants to see displayed. The echo magnitude values for an image
frame are therefore applied to a conventional scan converter 122, which
converts the stored image into a display format that is suitable for use
in driving a display device 124. The display device 124 typically includes
a conventional display driver 125 and a screen 126 (for example, LED or
CRT) that is divided into an X-Y (or polar) matrix or pattern of picture
elements or "pixels" that make up an image that the user can view and
interpret.
The image is displayed as a pattern of image elements that correspond to
the received echo magnitude from corresponding portions of one 2-D frame
of data from the interrogation region. Note that a displayed image element
will often be made up of more than one pixel, but that this will depend on
the relative resolutions of the scan and of the display. The invention
does not require any particular relative resolution.
Ultrasonic imaging may be done in any of several modes. One common mode is
the brightness or "B" mode, in which the display is typically gray-tone,
and the displayed intensity of each pixel corresponds to the amplitude of
the echo signal from a corresponding element or portion of the
interrogation region. In other words, the stronger the acoustic echo is
from a portion of the scanned region, the more brightly it is displayed.
Note that it is also possible to display intensity data using
"pseudo-colors," that is, such that different intensities (or intensity
intervals) are displayed using different assigned colors. For example,
increasing intensity can be displayed as increasingly more red.
The invention also includes a feature identification and measurement (ID&M)
sub-system 128, which is connected to or receives positional signals from
the cursor control device (preferably, trackball) 101, the display driver
125, and the memory 118 in order to access the 2-D frame data 120.
Connection with other system components such as the display driver and
memory may be either direct or indirect via some other component such as
the general processing system 102 or a dedicated intermediate circuit. The
ID&M sub-system may be a separate processor or cooperating group of
processors; alternatively, it may be combined with or incorporated in some
other processors in the system.
The general method according to the invention includes the following main
steps:
1) The user selects an image for display on the screen 126.
2) The user marks reference points on the display depending on a
user-specified assumption about the general geometry of the body structure
of interest.
3) The system applies a series of conversions, filters, and other
procedures to the image, identifies the structure, and automatically
measures the parameter of interest, which will normally be the path length
of an approximate line segment (for example, for a femur) or a measure of
circumference, diameter, or even area of a closed region such as a
cross-sectional display of the skull.
These steps are explained in detail below.
Image Selection
During the normal course of an ultrasound scan, the user will view the
display screen 126, which will display the series of frames corresponding
to the scan. When the body structure of interest is visible on the screen,
the user then directs the system to "freeze" the displayed data frame; the
user may do this in any known manner, such as pushing or releasing a
button, switch or key on the probe 106 or input device 100. Alternatively,
using known methods, the user may call up a pre-stored frame of image data
from an earlier scan. It is also possible according to the invention to
allow the system to generate continuously updated measurements, with no
need to stop the scan to "freeze" a frame, as long as sufficiently fast
processors are included in the system.
Reference Point Marking
It is known to display a cursor on a screen, such as the display screen
126. According to the invention, the user first maneuvers the input device
(preferably, a mouse or the trackball 101) to move a cursor on the display
screen until it points to or lies on a first reference point on the
displayed structure of interest. She then activates a button, key or other
known switching device to signal to the processing system 102 that the
corresponding first reference image element is at the first reference
point, and the processing system, via the display driver, then displays a
suitable mark, such as an "X", cross-hairs, etc., at this point. The user
then repeats this process to select and have marked a second reference
point on the displayed image, corresponding to a second reference image
element. Note that the scale of the image is known to the system and is
also normally displayed along one edge of the screen. It is consequently
possible for the system to apply known methods to calculate the linear
distance between any two given points on the display screen.
The preferred criterion for selecting reference points depends on the
assumed general geometry of the body structure of interest; the various
criteria are described in detail below. One should note, however, that it
is not necessary for the user to define the entire structure to be
measured by marking dots all along its path. Rather, in the preferred
embodiment of the invention, the user needs to designate at most two
delimiting reference points; indeed, in certain embodiments, the system
according to the invention can operate fully automatically and determine
the length parameter of interest with no user-input of reference points at
all. The system according to the invention then automatically determines
the remaining image points necessary to calculate the parameter of
interest, and performs the calculation. Since the processing systems 102
and 128 may operate many orders of magnitude faster than can a human
operator who is "clicking" on a large number of image points, the
invention greatly speeds up the process of measuring the displayed body
structure; furthermore, it produces more consistent and unbiased
measurement results than what a human operator can.
According to the invention, it is preferably assumed that the body
structures of interest will have either of two general geometries: mainly
closed and round or mainly open and linear. Measurements of generally
closed, round structures would include measurements of head circumference
(HC) and abdominal circumference (AC). According to the invention, other
features such as biparietal diameter (BPD) are essentially linear, but
characterize a generally closed, round structure and are determined using
procedures for such round structures. One example of a measurement of a
generally open and linear structure would be the measurement of femur
length (FL).
Note that body structures such as the head and femur will seldom if ever be
perfectly "round" or "straight," respectively. The invention does not
require them to be and, indeed, can in most normal cases even "fill in"
gaps in the image as long as the general shape is known.
The user may specify the assumed general geometry in any of several
different ways. Examples include keyboard entry and cursor-controlled
selection from a displayed menu, pull-down menu, or icon group giving the
selection for the possible general geometries. The geometry selections
could be words describing the general shape, such as "STRAIGHT," "ROUND,"
or "LINE," "CIRCLE," etc.; of the structure itself, such as "FL," "HC,"
"BPD," etc.; or, especially in the case of an icon group, even symbolic
choices such as small pictures of line segments, circles (for
circumference), diameters of circles, or shaded circles (for area
calculations).
In a conventional ultrasound scan, the areas of the interrogation region
with the strongest return signals are usually displayed brighter (closer
to the white end of the gray scale) than areas with weaker return signals;
bright areas typically correspond to body structures, since structural
boundaries tend to have relatively large changes in acoustic impedance.
Most of the image looks "dark." To make it easier to see features against
the white background of the drawings, this shading scheme is reversed in
those drawings that illustrate scan images, so that areas with stronger
return signals are shown darker.
Round Structures
The two most common closed structures of interest in obstetric ultrasound
examinations are the head and abdomen of the fetus and the parameters of
greatest interest are the circumference (HC and AC) and some diameter (in
the case of the head, BPD and OFD). The invention measures such mainly
closed structures in substantially the same way, although, as is described
further below, it is also able to make use of additional known structural
features of the fetal brain to improve the ability to identify and measure
the skull.
The main steps the invention follows for measuring closed structures are as
follows:
1) The assumed image of the structure is delimited to a portion of
interest.
2) The delimited portion is converted from the raster form in which it is
normally displayed into a polar form for analysis.
3) After optional but preferred sub-steps such as contrast enhancement and
weighting, the polar image is binarized so that all image elements are
preferably rendered as either fully "white" or fully "black."
4) The binarized image is filtered morphologically to further isolate the
image elements that correspond to the closed structure.
5) Curve boundaries are identified, filtered, and filled in as needed to
form a filtered representation of the structure.
6) An optimal approximating boundary function is determined and displayed
for the filtered representation, and its length (corresponding to the
circumference) is calculated and displayed.
7) If the length parameter of interest is a diameter, such as AD, BPD, or
OFD, then this is determined by evaluating the boundary function.
These steps are explained further below.
FIG. 2 illustrates an image of an ultrasound scan of a cross-section of the
head of a fetus. The skull appears as a generally elliptical closed
region, which, because of noise, deflection, and other acoustic properties
of the interrogation region, may have "breaks," for example, often where
the surface is parallel to the direction of propagation of the ultrasonic
scanning signals. The image will often also have visible returns from
relatively structured regions, which themselves have a pronounced curved
or linear shape. These might, for example, be returns from the mother's
own muscle tissue or uterine wall 202, or from fat layers 204. Other
visible returns may appear generally unstructured, such as the region
labelled 206 in the figure. All such returns are irrelevant (they are
noise) to measuring any distance parameter of the head and their influence
should therefore be eliminated or at least reduced; the way in which the
invention does this is described below.
Examinations of the head usually also have a visible return from the
mid-line 208, that is, the substantially linear region between the two
hemispheres of the brain. Although the invention is able to determine head
circumference and different diameters without mid-line information, the
preferred embodiment isolates the mid-line image and uses the
corresponding image portion to improve its ability to identify and measure
diameters.
Notice that most structured noise is located outside the generally
elliptical curve of the skull, whereas the midline is located inside the
curve. Notice that the skull usually more closely approximates an ellipse
than a circle, and that it may "bulge" more at the rear than at the front.
The way in which the invention uses these properties to advantage is
described further below.
FIG. 3 illustrates the same scan as FIG. 2, but shows certain
system-generated display features such as the references points 210, 212
(indicated as small crosses "+"), which the user selects in the manner
described above, as well as an OFD line 214 and a BPD line 216. As is well
known, the OFD line lies on or very close to the mid-line 208. The
invention preferably also generates and displays a line of circumference,
which shows the circumference that the invention determined based on the
image and used in measuring circumferential or diametral length. This line
is preferably superimposed on the display, but is not drawn in FIG. 2 to
avoid confusion with the skull image 200. The lines 214, 216 and the
circumference line are preferably displayed in a non-gray scale color so
that the user can see clearly where and what they are.
For head or abdominal measurements, the user should preferably mark as
reference points 210, 212 the approximate endpoints of what she estimates
to be the major axis (greatest diameter) of the curve 200. As is explained
below, this aids the invention not only by identifying two points assumed
to lie on or very near the curve 200, but also by setting a rough upward
bound on the diameter of the curve. The user could, however, also be
instructed to mark the assumed endpoint of the minor axis of the curve,
which would set a rough lower bound on the size of the curve. Some other
pair of reference points, preferably diametrically opposing, could also be
marked, but such a choice will in most cases not give as useful a starting
"guess" to the system. Furthermore, according to one alternative
embodiment of the invention, the system can isolate the curve 200 based on
only a single point (preferably near the center of the curve).
Any conventional coordinate system and scale may be used according to the
invention to define, both quantitatively and qualitatively, such terms as
"inside," "outside," as well as distances. The position of any point in
the interrogation region is therefore well-defined in known system
coordinates.
Structure Delimitation
In the preferred embodiment of the invention, the curve 200 is delimited by
an outer circle 300, whose radius is at least as large as the largest
possible radius of the curve 200, and an inner circle 302, whose radius is
at most as large as the smallest possible radius of the curve 200. There
are several ways according to the invention to determine the radii of the
delimiting circles 300 and 302.
In the preferred embodiment, in which the user is instructed to choose the
reference points 210, 212 to be the endpoints of the major axis of the
curve 200, the invention first designates an assumed center point at the
midpoint between the two reference points 210, 212. The distance from the
midpoint to either reference point is then the reference radius r.sub.ref.
The radii of the outer and inner circles can then be set equal to
predetermined percentages greater than and less than, respectively,
r.sub.ref. The percentages will depend on the assumed maximum eccentricity
of a head (or abdomen or other generally round structure of interest),
which can be determined by experiment. Alternatively the system may
include and use a pre-stored table of known, maximum outer radii (for
example, OFD for the head) for a fetus at any given gestational stage. The
user may then enter the approximate gestational stage, for example, in
weeks, before the system begins the measurement procedure.
As one alternative, it will often be adequate simply to set the radius of
the outer circle equal to r.sub.ref plus some predetermined small margin,
that is, to let the circle pass just on the outside of the reference
points. Rather than using percentages, the radius of the inner and outer
circles may alternatively be set to a distance corresponding to an
experimentally predetermined number of pixel values inside and outside the
references marks, measured along the line 214.
It is also possible, however, not to require or rely wholly on such prior
knowledge of eccentricity. Instead, the invention may divide the entire
image region into several angular sectors, and then divide each angular
sector into several concentric, radial tracks. In order to reduce the size
of the irrelevant area about the midline 208, the sectors may have a
minimum radial boundary set to an experimentally predetermined percentage
of the major radial distance. Alternatively, the system can calculate the
minimum radial boundary to be greater than half the length of the midline
208 (see FIG. 1), which may be identified and measured using a routine
described below. To avoid the possibility that this value is too large
(greater than the possible BPD), an upper limit for the minimum radius may
be set as a percentage of the distance r.sub.ref.
Related to this alternative implementation, the number of sectors and the
radial width of the tracks may be chosen by experiment. The average
intensity of each track is then calculated and the radius of the innermost
peak average intensity for each sector is identified. The maximum and
minimum peak radii are then also identified. The radius r.sub.min of the
inner circle 302 can then be set to a value that is less, by a preset
percentage, than the smallest "peak" radius. Similarly, the radius
r.sub.max of the outer circle 300 can be set to a value that is greater,
by a preset percentage, than the greatest "peak" radius. The greatest
innermost radius should be approximately equal to an experimentally
predetermined percentage of the reference radius r.sub.ref. Furthermore,
for heads, the radius to the innermost peak should be for the sector that
extends roughly perpendicular to the line 214. If either of these
assumptions is violated, then the system may apply default radius values
determined as above based on percentages of r.sub.ref.
Observe that delimiting the structure not only speeds up calculation times
but also, usually, "automatically" cuts out much noise.
Raster-to-Polar Conversion
As is usual, the image that the system displays to the user is in the
substantially Cartesian, raster format illustrated in FIGS. 2 and 3. This
is natural, since it maintains the scale and shape of the actual body
structures being imaged, assuming appropriate conventional beamforming and
scan conversion are provided. For purposes of structure identification and
measurement, however, the invention preferably converts the raster image
into polar form, with the calculated center point (the midpoint of the
line connecting the reference points) as the origin of the r-.theta.
(radius-angle) polar coordinate system. It is not necessary to display the
conversion to the user; rather, the intensity values of the raster scan
are stored in polar form in the memory.
FIG. 4 illustrates the image of FIGS. 2 and 3 in polar form. Since it is
known that the curve 200 lies completely outside of the inner circle 302
and inside the outer circle 300, only this annular region is preferably
converted and stored. With the chosen origin, the inner and outer circles
will map to straight lines and are shown as such in FIG. 4. The curve 200
will map to a wavy, substantially sinusoidal line; the waviness of the
line increases the more the curve 200 deviates from being a circle. The
region between the outer and inner delimiting circles 300, 302 thus
defines an annular search region for the image.
One advantage of setting r.sub.min and r.sub.max according to a number of
pixels offset from the reference points is that the circle through the
reference points will then map to a vertical line that divides the polar
representation into halves of equal width. This is therefore preferred,
although it is not necessary as long as the inner circle is chosen small
enough to certainly lie fully within the curve 200. FIG. 4 is drawn to
illustrate this.
It is not necessary to convert to polar representation every image element
between the delimiting circles 300, 302, although this may be done if the
necessary computations can be done if the additional time required to do
the calculations is acceptable in a given application. Rather, the polar
image illustrated in FIG. 4 may be compiled using the pixel intensity
information only along a number of radial rays that extend between the
delimiting circles. For example, assuming that the outer and inner circles
300, 302 are positioned m pixels beyond and within, respectively, the
reference mark, and n rays are spaced evenly over the 360.degree. extent
of the annular search region, then the annular search region will map to
an m-by-n pixel rectangular strip as shown in FIG. 4.
The more rays that are used, the greater will be the resolution polar
representation, but the longer it will take to perform the measurement.
The number of rays will therefore be determined by normal experimentation
given knowledge of the processing speed available in any given
application. In one prototype of the invention, 256 rays of 128 pixels in
length were evaluated and convened to polar form. In FIG. 4, the r-axis
would therefore represent a pixel width of 128 from r.sub.min to r.sub.max
and the .theta.-axis would represent 256 horizontal "strips" one pixel
wide, 128 pixels long, and with an angular spacing of approximately
360/256 degrees.
Image Binarization
In order to measure the curved body structure, the invention must first
determine which of the pixels in the image represent the structure. Alter
the structure has been delimited as described above, the elements in its
image still have intensity values throughout the gray-scale range of the
display. The invention preferably binarizes the image before further
filtering and measuring.
The simplest way to binarize the image of FIG. 4 is to determine by
experimentation and observation a threshold intensity value I.sub.t ; one
can then set to full bright ("1") each image element whose intensity value
is greater than I.sub.t and set all other element values to full dark
("0"). This will completely eliminate from consideration all noise below
I.sub.t, but in general it will be difficult to determine an absolute
value for the threshold I.sub.t that will be suitable for different images
or structures. For example, if an image is relatively dark (a low mean
brightness), then it may be set completely to black, even though the human
user herself might be able to discern the body structure in the weak
image.
One improvement the invention includes is that it chooses I.sub.t to be a
function of a maximum intensity value in at least a local portion of the
search region. It then compares a filtered functional value of the element
intensity values with I.sub.t and then sets them to full bright or full
dark accordingly. The preferred ways to determine I.sub.t and to filter
the image intensity values are described below.
Contrast Improvement
The first step in the binarization method in the preferred embodiment of
the invention is to increase the contrast of the polar image, which is
shown in FIG. 4. Common to all methods for improving contrast according to
the invention is that a turning point brightness is determined. A contrast
function is applied to the pixels in the polar image with the result that
elements whose intensity is greater than the turning point brightness are
made even brighter and elements whose intensity is less than the turning
point brightness are made even darker.
One way to increase contrast is by using a single-parameter contrast
function such as I.sub.cont =I.sub.in.sup..gamma. where I.sub.cont is the
intensity of a pixel after contrast improvement, I.sub.in is the input
intensity I.sub.in, and .gamma. is an experimentally determined parameter
that defines the turning point brightness. Since 0.ltoreq.I.sub.in
.ltoreq.1 (in certain cases, after standard normalization), then
0.ltoreq.I.sub.cont .ltoreq.1.
Contrast functions of two or more parameters may also be used. One example
is a sigmoid contrast function such as:
##EQU1##
where x=I.sub.in ; a is the turning point; and b determines the degree of
"stretching" of the intensity values about a. The value a may, for
example, be chosen equal to the average intensity of pixels in a
predetermined region and b may be set equal to the standard deviation of
intensity values for pixels over the same or over some other region. The
preferred regions over which the average and standard deviation are
determined are described below.
In FIG. 4, a constant angle image "strip" is labelled 400. The angular
width .delta..theta. of the strip may be any number of pixels, but is
preferably one pixel, so that the strip corresponds to a radial ray. In
the illustrated example, the strip extends only to the line through the
reference points at radius f.sub.ref ; this makes use of the fact that,
for heads, the most useful information about the curve 200 lies in the
left half plane of the polar plot, whereas the right half will typically
have a much lower signal-to-noise ratio. The strip could, however, extend
further, even to the outer circle 300, and preferably does so in the case
of imaging of an abdomen.
For the exponential contrast function, the value of a used for the pixels
in any given strip is the average intensity of the pixels in that strip.
The value of b, however, is preferably a function of the standard
deviation of intensity for all pixels in the left half plane (all pixels
from r.sub.min to r.sub.ref). The advantage of using a local mean a but a
global stretching factor b standard deviation is that portions of the
search region that have relatively low intensity more because of their
position, for example on the side of the head away from the transducer,
will not be darkened because of their position alone. The degree of
"stretching," however, will be determined by the same parameter b for all
pixels. Changes in contrast will therefore depend on relative brightness
rather than on position.
The parameter b may differ depending on the type of examination, but will
often be more or less constant for any given type. It is therefore
possible according to the invention to determine these contrast
"stretching" values for, for example, heads, livers, thyroids, or other
structures. The system can then save calculation effort simply by using
the appropriate prestored value.
Note that the values a may be determined based on only part of a radial
sector, for example, the left half-plane strip 400 in FIG. 4. This value,
however, is used in the contrast function that is applied to all pixels
over the full r.sub.min to r.sub.max width of the corresponding
.delta..theta. strip.
The abdomen normally doesn't have as many bright structures as the head,
since the structures for the mother and fetus are roughly the same and
dark regions are mostly amniotic fluid. Instead of a half-plane strip 400
as is illustrated in FIG. 4, it is preferred to evaluate the local
parameter a over the entire radial strip from r.sub.min to r.sub.max, or
within an annular sector centered on the r.sub.ref line but less than the
full width of the plot. The inventors have determined that two ways of
choosing a that produce good results for abdominal measurements are:
a=1/2.multidot..mu..sup.2
where .mu. is the average intensity value of the strip,
0.ltoreq..mu..ltoreq.1.
and
a =max[(1/2.multidot..mu..sup.2), (.mu.-.sigma..sub.cent), I.sub.min ]
where .sigma..sub.cent is the standard deviation of intensity values within
an annular strip centered on the r.sub.ref line extending, for example,
half way to r.sub.max and r.sub.min on either side and I.sub.min is the
minimum intensity value in the corresponding strip. For certain abdominal
images, the first term (1/2.multidot..mu..sup.2) can become very small and
the value .sigma..sub.cent can become large. Although these terms provide
good "stretching," that is, contrast improvement, they may occasionally
provide too low a turning point to be useful. Including I.sub.min thus
avoids having all or most pixels in a strip being set to "bright" ("1") in
such cases.
For other body structures, different turning point parameters a and
"stretching" parameters b may be determined by experiment. Indeed, other
contrast functions may be chosen if experience with imaging a particular
body structure indicates some advantageous function choice.
In order to avoid suspiciously rapid changes in the global parameter values
b from one frame of measurement to the next, it is also possible to set
this value equal to a weighted average of the current and one or more most
recent values. For example, the system could apply as .sigma. the value
.sigma.=.alpha..multidot..sigma..sub.new
+(1-.alpha.).multidot..sigma..sub.old, where .alpha. is chosen by
experiment. Another method for smoothing these parameters is to include
more lines (angles) in the neighborhood used for calculating the
parameters such as .mu. and .sigma.; moreover, smoothing even over such
multi-angle regions may be combined with previous values using a
time-decay factor such as the one described above.
Radial Weighting
Once the contrast function has been applied to all of the pixels of
interest in the search region, their intensity values are preferably
weighted such that the intensity value of a pixel is lower the farther it
is from the r.sub.ref line. Note that this is spatial weighting or
filtering, as opposed to purely brightness-derived weighting or filtering
used in the contrast-improvement step above.
In the preferred embodiment of the invention, a Gaussian weighting function
is applied over each radial strip. This is done by multiplying each
intensity value by the weighting factor:
##EQU2##
where r is the radial distance of the pixel from the center point and s is
an experimentally determined roll-off factor. Note that the pixel on the
r.sub.ref line retain their intensity values whereas pixels at the edges
of the search region (from r.sub.min to r.sub.max) are attenuated.
Other weighting functions may of course be used, such as triangular,
trigonometric, or parabolic windows. Furthermore, the weighting
calculations may be combined with those for contrast improvement.
Binarization Threshold
After the preferred but optional steps of contrast improvement and
weighting, the image is still in a gray-tone format, that is, the pixel
intensity values are distributed over a range of brightness. The final
step in binarizing the image is to select the threshold intensity value
I.sub.t and apply the threshold to the pixel intensity values so that the
remaining image consists of pixels whose values are either full bright
("1") or full dark ("0").
One way to select I.sub.t is as a global value. For example, the system may
evaluate the pixel intensity values for all pixels in the search region to
determine the maximum intensity I.sub.max. All pixels whose intensity
value is greater than or equal to an experimentally predetermined
percentage of I.sub.max, for example, 0.7.multidot.I.sub.max are then set
to "1" and all whose values are less than this value are set to "0".
In order t | | |
