BACKGROUND
Pneumoconiosis is a lung disease caused by long-term inhalation of
operative dust, such as coal, asbestos, silica, etc. For diagnosing whether a person is
suffering from pneumoconiosis and determining which stage of pneumoconiosis the
patient is experiencing, doctors typically need to compare chest X-ray photographic
films of the diagnosed person with some standard pneumoconiosis chest X-ray
photographic film samples according to diagnostic criteria of pneumoconiosis
established by some groups such as International Labor Office (ILO). When the chest
X-ray photographic film of the diagnosed person is determined to match one of the
standard pneumoconiosis chest X-ray photographic film samples according to the
doctors' knowledge and experience, the person is diagnosed that helshe is in the
defined stage corresponding to the matched standard pneumoconiosis chest X-ray
photographic film sample.
However, doctors must be trained by special professional organizations
and achieve corresponding professional qualification for diagnosing a patient with
pneumoconiosis. In addition, the chest X-ray photographic films of the person need
to satisfy special requirements, such as parameters of the X-ray generator, parameters
of the X-ray photographic films, etc. Since the diagnosed results are determined by
different doctors, different diagnosis results may be given based on the knowledge
and experience of each of the doctors, and manual diagnosis is time-consuming and
costly.
BRIEF DESCRIPTION
A diagnosis system and method of diagnosing a lung disease are
provided. The computer-aided diagnosis system includes first and second training
units, a segmentation unit, and an evaluation unit. The first training unit generates a
lung shape model based on first normal digital radiography (DR) chest X-ray image
samples. The second training unit generates at least one zone classifier model for
classifying abnormalities of the lung disease based on second normal DR chest X-ray
image samples and DR chest X-ray image samples people diagnosed with the lung
disease. The segmentation unit segments out a lung field image from a DR chest Xray
image to be diagnosed according to the lung shape model. The evaluation unit
evaluates the lung field image according to the at least one zone classifier model to
determine whether the DR chest X-ray image has abnormalities of the lung disease.
DRAWINGS
These and other features and aspects of embodiments of the present
invention will become better understood when the following detailed description is
read with reference to the accompanying drawings in which like characters represent
like parts throughout the drawings, wherein:
FIG. 1 is a block diagram of a computer-aided diagnosis system
according to one embodiment.
FIG. 2 is a block diagram of a first training unit of the system of FIG. 1
according to one embodiment.
FIG. 3 is a schematic view of an exemplary process of the first training
unit of FIG. 2 according to one embodiment.
FIG. 4 is a flowchart of an exemplary process of the first training unit of
FIG. 2 according to one embodiment.
FIG. 5 is a block diagram of a second training unit of the system of FIG.
1 according to one embodiment.
FIG. 6 is a flowchart of an exemplary process of the second training unit
of FIG. 5 according to one embodiment.
FIG. 7 is a schematic view of a right lung field and a left lung field of a
digital PA chest X-ray image.
FIG. 8 is a block diagram of a segmentation unit of the system of FIG. 1
according to one embodiment.
FIGs. 9-10 are schematic views of an exemplary process of the
segmentation unit of FIG. 8 according to one embodiment.
FIG. 11 is a flowchart of an exemplary process of the segmentation unit
of FIG. 8 according to one embodiment.
FIGs. 12 and 13 are schematic views of an exemplary process of a subdivision
unit of the system of FIG. 1 according to one embodiment.
DETAILED DESCRIPTION
Embodiments of the invention relate to a computer-aided diagnosis
system for diagnosing a lung disease. The computer-aided diagnosis system includes
first and second training units, a segmentation unit, and an evaluation unit. The first
training unit generates an average lung shape model based on a plurality of first
normal digital radiography (DR) chest X-ray image samples. The second training unit
generates at least one zone classifier model for classifying abnormalities of the lung
disease based on a plurality of second normal DR chest X-ray image samples and a
plurality of abnormal DR chest X-ray image samples representing the lung disease.
The segmentation unit segments a lung field image from a DR chest X-ray image to
be diagnosed according to the average lung shape model. The evaluation unit
evaluates the lung field image according to the at least one zone classifier model to
determine whether the DR chest X-ray image has abnormalities of the lung disease.
Unless defined otherwise, technical and scientific terms used herein have
the same meaning as is commonly understood by one of ordinary skill. The terms
"first", "second", and the like, as used herein do not denote any order, quantity, or
importance, but rather are used to distinguish one element from another. Also, the
terms "a" and "an" do not denote a limitation of quantity, but rather denote the
presence of at least one of the referenced items, and terms such as "front", "back",
"bottom", andor "top", unless otherwise noted, are merely used for convenience of
description, and are not limited to any one position or spatial orientation. Moreover,
the terms "coupled" and "connected" are not intended to distinguish between a direct
or indirect coupling/connection between two components. Rather, such components
may be directly or indirectly coupled/connected unless otherwise indicated.
Referring to FIG. 1, an exemplary computer-aided diagnosis system 100
for diagnosing pneumoconiosis is shown. The computer-aided diagnosis system 100
is used to directly process DR imageslradiographs (such as digital posterior-anterior
(PA) chest X-ray images) and diagnose pneumoconiosis according to processed DR
chest X-ray images. In one embodiment, the computer-aided diagnosis system 100
may be used as either a stand-alone tool (such as a computer) or embedded in a DR
system, or embedded in a picture archiving and communication system (PACS). Also,
the computer-aided diagnosis system 100 may be used in either real-time mode (i.e.
making diagnosis decision while a digital PA chest X-ray image is taken) or off-line
mode (i.e, making diagnosis decision on one or a batch of chest digital PA X-ray
images that are collected and stored in a computer media in advance).
In the embodiment of FIG. 1, the computer-aided diagnosis system 100
includes an image input unit 110, a first training unit 120, a segmentation unit 130, a
sub-division unit 140, a second training unit 150, a first classification unit 160, a
second classification unit 170, an evaluation unit 180, and a report output unit 190.
The computer-aided diagnosis system 100 may be programmed with software
instructions stored in a non-transitory computer-readable medium, which, when
executed by a processor, perform various operations of the system 100. The
computer-readable medium may include volatile and nonvolatile, removable and nonremovable
media implemented in any method or technology. The computer-readable
medium includes, but is not limited to, RAM, ROM, EEPROM, flash memory, digital
signal processor (DSP) or other memory technology, CD-ROM, digital versatile disks
(DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk
storage or other magnetic storage devices, or any other non-transitory medium which
can be used to store the desired information and which can be accessed by an
instruction execution system. The computer-aided diagnosis system 100 also may be
implemented by hardware or by combination of hardware and software.
The image input unit 110 is used to receive DR images to be processed,
and in particular, the DR images may include DR images to be trained and DR images
to be diagnosed. The DR images may be received from a DR system, a PACS, or a
data storing device. The report output unit 190 is used to output diagnosis results.
For example, the output diagnosis results can be displayed on an external monitor, or
stored in a database which may act as subsequent analysis data. In one embodiment,
the image input unit 110, the first training unit 120, the segmentation unit 130, the
sub-division unit 140, the second training unit 150, the first classification unit 160, the
second classification unit 170, the evaluation unit 180, and the report output unit 190
may be situated in discrete units andlor algorithms. In other embodiments, two or
more of these units of the system 100 may be integrated together in a common unit
and/or algorithm.
Referring to FIG. 2, an exemplary block diagram of the first training unit
120 is shown. The first training unit 120 is used to build training models for the
segmentation unit 130. In this embodiment, the first training unit 120 includes a
training image collecting module 121, a lung field labeling module 122, an aligning
module 123, a point distribution model (PDM) generation module 124, and an
average lung shape model generation module 125. In this embodiment, the first
training unit 120 together with the segmentation unit 130 applies an active shape
model (ASM) approach to extract an accurate lung field mask based on many digital
PA chest X-ray image samples collected by the training image collecting module 121
from the image input unit 110. In other embodiments, the first training unit 120
together with the segmentation unit 130 may apply other approaches to implement
similar functions, such as applying a level set approach, a rule-based reasoning
approach, a pixel classification approach, etc. The digital PA chest X-ray image
samples may include a number of digital PA chest X-ray images, such as twenty
normal persons' digital PA chest X-ray images, for example, and these digital images
may be processed through the type of ground truth (like binary digitals "0" and "1")
or other types.
Referring to FIGS. 3 and 4, a schematic view and a flowchart of an
exemplary process performed by the first training unit 120 are respectively shown. In
FIG. 3, one of a number of digital PA chest X-ray image samples 3 10 is labeled for
defining a lung field contour 301 by the lung field labeling module 122. In this
embodiment, the lung field contour 301 is defined by manually labeling many key
landmark points 302 along the lung field contour 301 corresponding to block 401 of
the flowchart of FIG. 4. The other chest X-ray image samples 310 are also labeled
according in a similar manner. For obtaining averageloptimal key landmark points
from all of chest X-ray image samples 3 10, these key landmark points 302 of all chest
X-ray image samples 3 10 are aligned by appropriate algorithms, such as by scaling,
rotating, and translating etc. corresponding to block 402 of the flowchart of FIG. 4
through the aligning module 123.
After aligning the key landmark points 302, a group of average/optimal
key landmark points 303 are calculated by the PDM generation module 124, namely
an averageloptimal PDM 320 is generated, corresponding to block 403 of the
flowchart of FIG. 4. After the PDM 320 is generated, an average lung shape model
330 is derived from the PDM 320 by smoothly connecting adjacent key landmark
points 303 of the PDM 320, corresponding to block 404 of the flowchart of FIG. 4
through the average lung shape model generation module 125. The average lung
shape model 330 may include two lung fields 304 which are characterized as two
averageloptima1 lung fields from the chest X-ray image samples 310 of different
people, such as twenty or any number of normal persons. In one embodiment, a shape
parameter, a pose parameter, and other related parameters are used to control the
deformation of the average lung shape model 330 are also calculated from the PDM
320 by appropriate algorithms.
Referring to FIG. 5, an exemplary block diagram of the second training
unit 150 is shown. The second training unit 150 is used to build training models for
the first classification unit 160 and the second classification unit 170. In this
embodiment, the second training unit 150 includes a training image collecting module
15 1, an image division module 152, a feature extraction module 153, a feature
selection module 154, and a classifier model building module 155. In this
embodiment, the second training unit 150 applies texture analysis algorithms to
extract features of pneumoconiosis abnormalities based on many digital PA chest Xray
image samples collected by the training image collecting module 151 from the
image input unit 110. For example, the texture analysis algorithms may include a
multi-scale filter bank algorithm, a histogram algorithm, a co-occurrence matrix
algorithm, and a spectral/spatial frequency domain algorithm, etc. In other
embodiments, the second training unit 150 may apply other appropriate approaches to
extract features of pneumoconiosis abnormalities. The digital PA chest X-ray image
samples may include digital PA chest X-ray images, such as 100 or any suitable
number of normal digital PA chest X-ray images and abnormal pneumoconiosis
digital PA chest X-ray images which have been accurately diagnosed by doctors in
advance and these digital images may be processed through the type of ground truth.
Referring to FIG. 6, a flowchart of an exemplary process of the second
training unit 150 is shown. At block 601, the digital PA chest X-ray image samples
are collected by the training image collecting module 151 from the image input unit
110. The number of the digital PA chest X-ray image samples can be adjusted
according to the updated database of pneumoconiosis.
At block 602, in this embodiment, each one of the image samples is
divided into six zones (namely regions of interest (ROI)) by the image division
module 152. Referring to FIG. 7, a digital PA chest X-ray image may include a right
lung field A and a left lung field B, and are divided into six zones R1, R2, R3, Ll, L2,
L3 evenly, for example. In other embodiments, the number of zones can be adjusted
according to different requirements. The image samples can be divided through
manual method like the method implemented in the lung field labeling module 122
mentioned before. The image samples also can be divided through automatic method
which will be particularly described in latter paragraphs related to the segmentation
unit 130 and the sub-division unit 140. In other embodiments, the image samples may
not be divided according to different requirements.
At block 603, the features of pneumoconiosis abnormalities of each zone
of each of these digital PA chest X-ray image samples are extracted by the feature
extraction module 153. In one embodiment, because abnormalities appeared in digital
PA chest X-ray image samples are of various size, of various degree of opacities,
rounded, or irregular, the image samples may be filtered (namely enhanced) in
advance by highlighting small opacities or improving image contrast, or other image
processing through a multi-scale filter bank algorithm, for example.
After filtering the image samples, the features of pneumoconiosis
abnormalities of each zone of each of the digital PA chest X-ray image samples are
easier to extract. In one embodiment, the significant features are calculated through a
histogram algorithm, a co-occurrence matrix algorithm, and a spectrallspatial
frequency domain algorithm, for example. Then, sufficient features of
pneumoconiosis abnormalities of each zone of each of the digital PA chest X-ray
image samples are extracted out, and may be recorded as one or more feature vectors
to represent texture features in each zone of each of the image samples. These
features may be characterized as mean, standard deviation (SD), skewness, kurtosis,
energy, entropy, homogeneity, correlation, inertia, horizontal filter, vertical filter,
root-mean-square (RMS), and first moment, for example.
At block 604, the most relevant features of the extracted features of
pneumoconiosis abnormalities are selected according to useful relevant characteristic
of pneumoconiosis by the feature selection module 154. The feature vectors may
include a lot of features (for example 300 features), these features may include some
irrelevant and redundant features (for example 250 irrelevant and redundant features),
and so the feature selection step can find the most relevant features (for example 50
most relevant features) and delete the irrelevant and redundant features which can
improve the performance of training models. Many standard data analysis algorithms
are utilized for feature selection. For example, in one embodiment, when some
features are determined to be very similar to each other, only one of them is retained
to represent their common characteristic, and the others are deleted. In other
embodiments, these irrelevant and redundant features also can be deleted by
implementing other selection algorithms, such as a correlation threshold algorithm, a
hierarchical clustering algorithm, a choice of feature algorithm, for example.
At block 605, six zone classifier models corresponding to the six zones
R1, R2, R3, L1, L2, and L3 are respectively built based on the selected features and
the diagnosed results of the image samples by doctors in advance through the
classifier model building module 155. The zone classifier models are also built based
on one kind of diagnostic criteria of pneumoconiosis, for example a version (GBZ 70-
2009) established by the Chinese government. In one embodiment, each zone
classifier model may include a feature vector corresponding to different
pneumoconiosis abnormalities. For example, one selected feature from the feature
vector may correspond to a pneumoconiosis abnormality such as a small opacity
having a diameter that is not greater than 3rnm.
The following table shows the exemplary stage definitions of
pneumoconiosis in the diagnostic criteria of pneumoconiosis of the Chinese
Government. This table is for exemplary purposes and does not limit the scope of
embodiments of the invention.
In the table, small opacities include rounded opacities and irregular opacities.
Rounded opacities include a "p" type rounded opacities whose diameter is not greater
than 1.5mm, "q" type rounded opacities whose diameter is between from 1.5mm to
3mm, and "r" type rounded opacities whose diameter is between from 3mm to 1Omm.
Irregular opacities include "s" type irregular opacities whose width is not greater than
1.5mm, "t" type irregular opacities whose width is between from 1.5mm to 3mm, and
"u" type irregular opacities whose width is between from 3mm to 1Omm. The
diameter or width of the large opacities is greater than 10mm. The profusion level of
small opacities includes four general levels which are 0, 1, 2, and 3, and the overall
profusion level represents the highest profusion level in all six zones. The profusion
levels of the affected zones are 1, 2, or 3. The pneumoconiosis includes three stages
which are I, 11, and 111. According to above table, the stage of a pneumoconiosis
image can be diagnosed.
Stage
I
I1
I11
Because the digital PA chest X-ray image samples are correctly
diagnosed by doctors in advance according to the diagnostic criteria of
pneumoconiosis, the zone classifier models can be built based on the diagnosed result
Opacities Size
Small Opacities
Small Opacities
Large Opacities
Small Opacities
Small and
Large Opacities
Overall Profusion Level
1
2
3
3
3
Affected ZonesISize
2 2 zones
> 4 zones
4 zones
Long diameter 1 20 mm;
Short diameter 1 10 mm
-
> 4 zones
> 4 zones
of the image samples corresponding to the selected features related to all relevant
pneumoconiosis characteristics. In one embodiment, the zone classifier models are
built through a support vector machine (SVM) algorithm. Each zone classifier model
can determine affected levels (namely the opacities size and the profusion level) of
the corresponding zone of an input digital PA chest X-ray image. In other
embodiments, the zone classifier models also can be built through other appropriate
classifier building algorithms. Otherwise, when the diagnostic criteria of
pneumoconiosis orland the image samples are updated for some reasons, the zone
classifier models can be rebuilt according to the updated diagnostic criteria of
pneumoconiosis andor the updated image samples. In other embodiments, the zone
classifier models also can be built from all of the extracted features, and thus the
feature selection module 154 is omitted accordingly. In other embodiments, if the
image samples are not divided, only one classifier model is built accordingly.
Referring to FIG. 1 again, after the training processes, an average lung
shape model is generated by the first training unit 120, and six zone classifier models
are generated by the second training unit 150, then the computer-aided diagnosis
system 100 can automatically diagnose a digital PA chest X-ray image to determine
whether it is a pneumoconiosis image andor determine which stage of a
pneumoconiosis image reaches. In one embodiment, the digital PA chest X-ray image
is received by the image input unit 110 and then transmitted to the segmentation unit
130 may be as ground truth type.
Refemng to FIG. 8 and FIG. 11, an exemplary block diagram of the
segmentation unit 130 and a flowchart of an exemplary process of the segmentation
unit 130 are respectively shown. The segmentation unit 130 is used to segment out a
lung field mask from the digital PA chest X-ray image for the sub-division unit 140.
In this embodiment, the segmentation unit 130 includes a lung position mask
extraction module 13 1, a lung thresholding image extraction module 132, an initial
lung contour extraction module 133, and a lung field mask generating module 134. In
this embodiment, the segmentation unit 130 applies an ostu thresholding algorithm, a
histogram equalization algorithm, and an ASM algorithm to segment out the lung
field mask based on the average lung shape model 330 generated by the first training
unit 120. In other embodiments, the segmentation unit 130 may apply other
appropriate approaches to segment out the lung field mask.
Referring to FIG. 9 and FIG. 1 1 together, FIG. 9 shows a schematic view
of an exemplary process of the lung position mask extraction module 13 1, the lung
thresholding image extraction module 132, and the initial lung contour extraction
module 133 corresponding to blocks 120 1, 1202, and 1203 of FIG. 1 1. At block 120 1,
a lung position mask 930 is extracted through appropriate algorithms by the lung
position mask extraction module 131. For example, a chest position mask 920 is
extracted from the digital PA chest X-ray image 910, and then the lung posi;ion mask
930 is extracted from the chest field mask 920 through appropriate image processes.
At block 1202, a lung thresholding image 950 is extracted through appropriate
algorithms by the lung thresholding image extraction module 132. For example, an
enhanced chest X-ray image 940 is processed from the digital PA chest X-ray image
910, and then the lung thresholding image 950 is extracted from the enhanced chest
X-ray image 940 through appropriate image processing. At block 1203, an initial
lung contour mask 960 is obtained by superimposing the lung position mask 930 and
the lung thresholding image 950 through the initial lung contour extraction module
133, for example.
Referring to FIG. 10 and FIG. 1 1 together, FIG. 10 shows a schematic
view of an exemplary process of the lung field mask generating module 134
corresponding to block 1204 of FIG. I I. At block 1204, a lung field mask 970 is
extracted based on the average lung shape model 330 and the initial lung contour
mask 960 by the lung field mask generating module 134. As described before, the
average lung shape model 330 represents two average lung fields from digital PA
chest X-ray image samples of different people, and the initial lung contour mask 960
represents a rough shape of the lung field of the digital PA chest X-ray image 910.
Therefore, the initial lung contour mask 960 can be transformed to the lung field mask
970 according to shape parameters of the average lung shape model 330. The lung
field mask 970 represents an accurate area of the lung field of the digital PA chest Xray
image 910.
Referring to FIGS. 12 and 13, two schematic views of representing an
exemplary process of the sub-division unit 140 are shown. For more accurately
diagnosing the stage of pneumoconiosis of the digital PA chest X-ray image 910, the
two lung fields 982 of the lung field mask 970 are divided into six zones rl, r2, r3,11,
12, and 13 (see FIG. 12) according to the diagnostic criteria of pneumoconiosis
mentioned above. If the diagnostic criteria of pneumoconiosis mentioned above are
updated for some reasons in the future, the two lung fields 972 of the lung field mask
970 may be divided accordingly, such as divided into twelve zones for example. In
some embodiments, the two lung fields 972 of the lung field mask 970 may not be
divided according to requirements.
Referring to FIG. 13, a lung field image 980 including two divided lung
fields 982 is extracted based on the lung field mask 970 and the digital PA chest Xray
image 910. For example, the divided lung field mask 970 is overlapped on the
digital PA chest X-ray image 910, and then the divided lung field mask 970 encloses a
field which is the lung field image 980 also divided into six zones rl, r2, r3,11, 12, and
13. The extracted lung field image 980 is the expected image to be processed for
subsequent diagnosing/evaluation processes in the evaluation unit 180. In one
embodiment, the expected image acts as ground truth type. In subsequent diagnosed
processes, only the digital data from each of the six zones rl, r2, r3, 11, 12, and 13 of
the extracted lung field image 980 are used to diagnose, which can improve accuracy,
sensitivity, and also can save time. In some embodiments, if the two lung fields 972
of the lung field mask 970 are not divided, then the sub-division unit 140 is omitted,
and an undivided lung field image is extracted based on the undivided lung field mask
970 and the digital PA chest X-ray image 910 by the segmentation unit 130
accordingly.
Referring to FIG. 1 again, the first classification unit 160 is used to build
a screening classifier model according to the six zone classifier models based on the
diagnostic criteria of pneumoconiosis mentioned above. The screening classifier
model can determine whether the digital PA chest X-ray image 910 is a
pneumoconiosis image. For example, when the digital PA chest X-ray image 910 has
a nonzero profusion level zone, it is determined that this digital PA chest X-ray image
910 is a pneumoconiosis image. When the computer-aided diagnosis system 100 is
only used to determine whether the digital PA chest X-ray image 910 is a
pneumoconiosis image, the first classification unit 160 is used and the second
classification unit 170 is omitted, which can save time.
The second classification unit 170 is used to build a stage classifier
model according to the six zone classifier models based on the diagnostic criteria of
pneumoconiosis mentioned above. The stage classifier model can determine which
stage the digital PA chest X-ray image 910 reaches. For example, if the number of
the affected zones of the digital PA chest X-ray image 910 is greater than four and the
overall profusion level is 2, and there are no large opacities in the image 910, then the
pneumoconiosis image 910 is diagnosed at stage 11. In other embodiments, the
second training unit 150, the first classification unit 160, and the second classification
unit 170 can be integrated together in a common unit andlor algorithm and can build a
whole classifier model to diagnose different disease results according to different
requirements.
Referring to FIG. 1 again, the evaluation unit 180 receives the digital
data of each of the six zones rl, r2, r3,11, 12, and 13 of the extracted lung field image
980 from the sub-division unit 140 and extracts the selected features from each of the
zones rl, r2, r3, 11, 12, and 13 of the extracted lung field image 980. The selected
features are determined by the feature selection module 154 of the second training
unit 150. After that, the evaluation unit 18 can diagnose the digital PA chest X-ray
image 910 based on the extracted the selected features with the screening classifier
model and the stage classifier model. One the one hand, the evaluation unit 180
determines whether the digital PA chest X-ray image 910 is a pneumoconiosis image
by comparing the extracted features from each of the six zones rl, r2, r3,11, 12, and 13
of the extracted lung field image 980 with the screening classifier model from the first
classification unit 160. On the other hand, the evaluation unit 180 determines which
stage the digital PA chest X-ray image 910 reaches by comparing the extracted
features from each of the six zones rl, r2, r3, 11, 12, and 13 of the extracted lung field
image 980 with the stage classifier model from the second classification unit 170.
In order to obtain a more accurate result, the evaluation unit 180 may
evaluate several times of the digital PA chest X-ray image 910 and get an optimal
result by analyzing different results. After that, the report output unit 190 will output
a diagnosed report which can be monitored by operators, such as doctors. In other
embodiments, if the evaluation unit 180 only evaluates whether the digital PA chest
X-ray image 910 has abnormalities of pneumoconiosis, the evaluation unit 180 can
only use the six zone classifier models to determine, and then the first classification
unit 160 and the second classification unit 170 are omitted accordingly. In other
embodiments, the computer-aided diagnosis system 100 also can diagnose other kinds
of lung diseases by training corresponding image samples and building corresponding
classifier models.
In one embodiment, the computer-aided diagnosis system 100 may be
operated by a graphic user interface (GUI), which can easily interact with operators.
The GUI may include button area used to execute functions of the units of the
computer-aided diagnosis system 100, image area used to show the digital PA chest
X-ray image, the extracted divided images, and other expected images, and report
display area used to display report results. For example, if it only needs to implement
screening function, a screening function button is switched on, and then the computeraided
diagnosis system 100 only implement screening function (namely only use the
first classification unit 160 but not use the second classification unit 170), which can
save time. Because the computer-aided diagnosis system 100 diagnoses
pneumoconiosis disease according to objective data analysis but not according to
subjective analysis, which can obtain a relatively objective and correct result.
Furthermore, the computer-aided diagnosis system 100 can save time and costs due to
only analyzing digital PA chest X-ray images through predetermined algorithms but
not analyzing chest X-ray photographic films through comparing with lots of standard
pneumoconiosis chest X-ray photographic film samples.
While exemplary embodiments of the invention have, it will be
understood by those skilled in the art that various changes may be made and
equivalents may be substituted for elements thereof without departing from the scope
of the invention. In addition, many modifications may be made to adapt a particular
situation or material to the teachings of the invention without departing from the
essential scope thereof. Therefore, it is intended that the invention not be limited to
the particular embodiments disclosed as the best mode contemplated for carrying out
this invention, but that the invention will include all embodiments falling within the
scope of the appended claims.

WE CLAIM:
1. A computer-aided diagnosis system for diagnosing a lung disease,
comprising:
a first training unit for generating an average lung shape model based on a
plurality of first normal digital radiography (DR) chest X-ray image samples;
a second training unit for generating at least one zone classifier model for
classifying abnormalities of the lung disease based on a plurality of second normal
DR chest X-ray image samples and a plurality of abnormal DR chest X-ray image
samples;
^ ^ a segmentation unit for segmenting a lung field image from a DR chest X-ray
image to be diagnosed according to the average lung shape model; and
an evaluation unit for evaluating the lung field image according to the at least
one zone classifier model to determine whether the DR chest X-ray image has
abnormalities of the lung disease.
2. The computer-aided diagnosis system of claim 1, fiarther comprising:
a first classification unit for building a screening classifier model according to
the at least one zone classifier model, the evaluation unit for evaluating the lung field
image according to the screening classifier model to determine whether the DR chest
X-ray image is a lung disease image.
3. The computer-aided diagnosis system of claim 1, further comprising:
a second classification unit for building a stage classifier model according to
the at least one zone classifier model, the evaluation unit for evaluating the lung field
18
image according to the stage classifier model to determine which stage the DR chest
X-ray image reaches.
4. The computer-aided diagnosis system of claim 1, further comprising:
a sub-division unit for dividing the lung field image into a plurality of zones.
5. The computer-aided diagnosis system of claim 1, wherein the first training
unit comprises:
^ a lung field labeling module for labeling a plurality of key landmark points
along a lung field contour of each of the plurality of first normal DR chest X-ray
image samples;
an aligning module for aligning the key landmark points of the plurality of
first normal DR chest X-ray image samples;
a point distribution model (PDM) generation module for generating a plurality
of average key landmark points as a PDM by calculating the aligned key landmark
points; and
an average lung shape model generation module for generating the average
lung shape model by smoothly connecting every adjacent key landmark points of the
PDM.
#
6. The computer-aided diagnosis system of claim 1, wherein the second
training unit comprises:
a feature extraction module for extracting features of abnormalities of the lung
disease based on the plurality of second normal DR chest X-ray image samples and
the plurality of abnormal DR chest X-ray image samples; and
19
a classifier model building module for generating the at least one zone
classifier model from the extracted features according to diagnosed results of the
plurality of second normal DR chest X-ray image samples and the plurality of
abnormal DR chest X-ray image samples corresponding to diagnostic criteria of the
lung disease.
7. The computer-aided diagnosis system of claim 6, wherein the second
training unit further comprises an image division module for dividing each of the
plurality of second normal DR chest X-ray image samples and the plurality of
abnormal DR chest X-ray image samples into a plurality of zones, wherein the at least
,.^ one zone classifier model comprises a plurality of classifier models corresponding to
^ ^ the plurality of zones.
8. The computer-aided diagnosis system of claim 6, wherein the second
training unit further comprises a feature selection module for selecting most relevant
features from the extracting features, the at least one zone classifier model is
generated from the selected features.
9. The computer-aided diagnosis system of claim 1, wherein the segmentation
unit comprises:
1^^ an initial lung contour extraction module for generating an initial lung contour
mask; and
a lung field mask generating module for generating a lung field mask by
transforming the initial lung contour mask according to shape parameters of the
average lung shape model, wherein the lung field image is segmented out by using the
lung field mask.
20
10. The computer-aided diagnosis system of claim 9, wherein the initial lung
contour mask is generated by extracting a lung position mask from the DR chest Xray
image, extracting a lung thresholding image from the DR chest X-ray image, and
superimposing the lung position mask and the limg thresholding image.
11. The computer-aided diagnosis system of claim 1, wherein the lung disease
comprises pneumoconiosis.
12. A computer-aided diagnosis method for diagnosing a lung disease,
^ ^ comprising:
generating an average lung shape model based on a plurality of first normal
digital radiography (DR) chest X-ray image samples;
generating at least one zone classifier model for classifying abnormalities of
the lung disease based on a plurality of second normal DR chest X-ray image samples
and a plurality of abnormal DR chest X-ray image samples;
segmenting a lung field image from a DR chest X-ray image to be diagnosed
according to the average lung shape model; and
evaluating the lung field image according to the at least one zone classifier
model to determine whether the DR chest X-ray image has abnormalities of the lung
disease.
13. The computer-aided diagnosis method of claim 12, fiirther comprising:
building a screening classifier model according to the at least one zone
classifier model, and evaluating the lung field image according to the screening
21
classifier model to determine whether the DR chest X-ray image is a lung disease
image.
14. The computer-aided diagnosis method of claim 12, further comprising:
building a stage classifier model according to the at least one zone classifier
model, and evaluating the lung field image according to the stage classifier model to
determine which stage the DR chest X-ray image reaches.
15. The computer-aided diagnosis method of claim 12, further comprising:
dividing the lung field image into a plurality of zones.
16. The computer-aided diagnosis method of claim 12, wherein generating an
average lung shape model based on a plurality of first normal digital radiography (DR)
chest X-ray image samples comprises:
labeling a plurality of key landmark points along a lung field contour of each
of the plurality of first normal DR chest X-ray image samples;
aligning the key landmark points of the plurality of first normal DR chest Xray
image samples;
I^P generating a plurality of average key landmark points as a point distribution
model (PDM) by calculating the aligned key landmark points; and
generating the average lung shape model by smoothly connecting every
adjacent key landmark points of the PDM.
22
17. The computer-aided diagnosis metiiod of claim 12, wherein the generating
at least one zone classifier model for classifying abnormalities of the lung disease
based on a plurality of second normal DR chest X-ray image samples and a plurality
of abnormal DR chest X-ray image samples comprises:
extracting features of abnormalities of the lung disease based on the plurality
of second normal DR chest X-ray image samples and the plurality of abnormal DR
chest X-ray image samples; and
generating the at least one zone classifier model from the extracted features
according to diagnosed results of the plurality of second normal DR chest X-ray
image samples and the plurality of abnormal DR chest X-ray image samples
^ corresponding to diagnostic criteria of the lung disease.
18. The computer-aided diagnosis method of claim 17, wherein generating at
least one zone classifier model for classifying abnormalities of the lung disease based
on a plurality of second normal DR chest X-ray image samples and a plurality of
abnormal DR chest X-ray image samples further comprises:
dividing each of the plurality of second normal DR chest X-ray image samples
and the plurality of abnormal DR chest X-ray image samples into a plurality of zones,
wherein the at least one zone classifier model comprises a plurality of classifier
models corresponding to the plurality of zones.
19. The computer-aided diagnosis method of claim 17, wherein generating at
least one zone classifier model for classifying abnormalities of the lung disease based
on a plurality of second normal DR chest X-ray image samples and a plurality of
abnormal DR chest X-ray image samples further comprises:
selecting most relevant features from the extracting features, the at least one
zone classifier model is generated from the selected features.
23
20. The computer-aided diagnosis method of claim 12, wherein segmenting
out a lung field image from a DR chest X-ray image to be diagnosed according to the
average lung shape model comprises:
extracting a lung position mask from the DR chest X-ray image;
extracting a lung thresholding image from the DR chest X-ray image
generating an initial lung contour mask by superimposing the lung position
mask and the lung thresholding image; and
generating a lung field mask by transforming the initial lung contour mask P according to shape parameters of the average lung shape model, wherein the lung field image is segmented out by using the lung field mask.