Computer-based objective quantitative assessment of
pulmonary parenchyma via X-ray CT
Renuka Uppaluri1,
Geoffrey McLennan2,
Milan Sonka3,
Eric A. Hoffman1,4
- 1Department of Radiology
University of Iowa, Iowa City, IA 52242
- 2Department of Internal Medicine
University of Iowa, Iowa City, IA 52242
- 3Department of Electrical and Computer Engineering
University of Iowa, Iowa City, IA 52242
- 4Department of Biomedical Engineering
University of Iowa, Iowa City, IA 52242
[ home
| search
| contact us
]
Table of Contents
This paper is a review of our recent studies using a texture-based tissue
characterization method called the Adaptive Multiple Feature Method. This computerized method is
automated and performs tissue classification based upon the training
acquired on a set of representative examples. The AMFM has been applied to
several different discrimination tasks including normal subjects, subjects with
interstitial lung disease, smokers, asbestos-exposed subjects, and subjects with
cystic fibrosis. The AMFM
has also been applied to data acquired using different scanners and scanning
protocols. The AMFM has shown to be successful and better
than other existing techniques in discriminating the tissues under consideration.
We demonstrate that the AMFM is considerably more sensitive and specific in
characterizing the lung, especially in the presence of mixed pathology, as compared
to more commonly used methods. Evidence is presented suggesting that the AMFM is highly
sensitive to some of the earliest disease processes.
Tissue characterization, texture, smoking, pulmonary, interstitial lung
disease
In recent years, it has become critical
that an objective quantitative lung assessment tool be developed which can be
used to measure outcomes for new drug therapies and surgical interventions. The
Adaptive Multiple Feature Method (AMFM) is such a computer-based technique for the
characterization of the pulmonary parenchyma from X-ray CT. While previously
reported methods using mean lung density information or the lowest fifth
percentile of the density histogram of the image have been shown to be reasonably successful for
characterizing emphysema, these methods have not been successful in identifying other
parenchymal diseases.
The AMFM has proven to be applicable
to a wide variety of tissue characterization problems relating to the
lung. This paper provides an overview of our recent successes with the AMFM
towards providing a user-independent quantitative assessment of lung. Specifically,
we will discuss the evolution of the AMFM as a method and its applications
to emphysema [1,2], interstitial lung diseases
[3,4],
smoking-related pathologies [5], asbestosis [6], and
cystic fibrosis [7].
The AMFM is a pattern recognition method which basically consists of three steps:
feature extraction, optimal feature selection, and classification. Prior to
feature extraction, pre-processing of the images was also performed and appropriate
regions of interest were selected.
When the AMFM was first designed, 17 texture features were extracted. These
were the grey level distribution features (mean, variance, skewness, kurtosis,
grey level entropy), run length features (short run emphasis, long run emphasis,
grey level non-uniformity, run length non-uniformity, run percentage), co-occurrence
matrix measures (angular second moment, inertia, entropy, contrast, correlation, inverse
difference moment), and a fractal measure (geometric fractal dimension).
In later studies, five other novel fractal measures were added. These were stochastic
fractal dimension (SFD) measures based on the fractional Brownian motion model; SFD
mean, SFD variance, SFD skewness, SFD kurtosis, SFD entropy. Still later, based
on the partial success of the lowest fifth percentile of the density histogram as
a measure of emphysema, five percentile-based measures were added. These were the
lowest fifth percentile of the histogram, highest fifth percentile of the histogram,
mean - lowest fifth percentile, highest fifth percentile
- mean, ratio of (highest fifth percentile - mean) over (mean - lowest fifth).
The optimal feature selection was performed using the ``divergence" measure
along with correlation analysis.
As a classifier, a minimum distance classifier was initially used for ease of implementation.
This was later replaced by a non-linear Bayesian classifier which performs
classifications based on a minimum loss criterion.
In all discrimination tasks, the data available was split arbitrarily into two independent sets - a
training and a test set. The training data was used to find an optimal subset of features
that best discriminated the samples in the training set. This optimal subset of features
was then extracted on the test set and classification was performed using the learning
acquired in the training stage. The success of the method in performing the discrimination
task was assessed by computing the accuracy (percent correctly classified samples of
all subject groups under consideration) or the classification rate (percent of correctly
classified sample of each subject group under consideration). Sensitivity and specificity
were computed in some studies.
Our first report of the partially designed AMFM and its applications can be
found in Uppaluri et al [1]. This served as a pilot study to assess the potential
of such a method for tissue characterization. The salient features of this work
are listed below:
- Data
Data included 10 normal and 10 emphysema
subjects. They were scanned either prone or supine with an Imatron C-150 XL electron
beam CT (EBCT) scanner (South San Francisco, CA). The collimation ranged from 3 mm to 10 mm.
Three slices of each subject were
used for the analysis.
- Method
There were 17 features used which were derived from the grey level distribution, run length,
co-occurrence matrix measures, and the geometric fractal dimension. A minimum distance classifier
was used following optimal feature selection.
- Findings
The first experiment was the global
analysis study where the goal was to differentiate the the whole CT slices of normal
and emphysematous subjects. It was found that these two subject groups could be discriminated
with 93.0% accuracy. In the second experiment, the sensitivity of the method was tested
by applying it for differentiating subtle differences in textures. Specifically,
the task was to discriminate between the anterior one-third vs. the posterior one-third
of a normal slice. Such differences were known to occur because of gravity. The accuracy in
differentiating these groups was 86.6%. In the third experiment, the lung on the CT slice
was divided in to 6 regions, anterior to posterior, and regional analysis
was performed. The AMFM was applied to differentiate
normal and emphysema samples derived from the same region (one of the six) on the slice.
This was performed with an accuracy of 82.9
5.7% over all the 6 regions.
It was concluded that computerized texture analysis was indeed applicable for tissue
characterization. However, this study had limitations. The slice thickness of the scans
varied from subject to subject.
Also, in the regional analysis experiment, any sample with at least 20.0% emphysema
(as assessed visually by an experienced observer)
was used. This probably led to the inclusion of emphysema samples having both ``normal"
and ``emphysema" regions in them. The lack of representative pure emphysema samples in the training
set probably led to ambiguous training and hence the accuracies of correct recognition were
lower.
In a following study [2], these limitations were overcome by data scanned using
a standardized
protocol. The AMFM was compared against previously reported methods for identifying
emphysema, namely, the mean lung density (MLD) method and the lowest fifth percentile of
the histogram (HIST) method.
- Data
Data included 9 normal and 10 emphysema subjects scanned with an Imatron C-150 XL EBCT
scanner. The normal subjects were scanned prone and the emphysema subjects were scanned
supine. Standardized protocol of same slice thickness (3 mm), similar field of view, and
same inspiration level (full inspiration) was used. Four slices from each subject were used
for the analysis.
- Method
As before, the grey level distribution, run length, co-occurrence
matrix measures, and the geometric fractal dimension were used as features.
A Bayesian classifier was used in place of the minimum distance classifier after the optimal
feature selection.
- Findings
The same three experiments as in the previous study were performed. In the global analysis,
the AMFM differentiated the normal from the emphysematous scans with 100.0% accuracy as
compared to accuracies of 94.7% and 97.4% by the MLD and HIST methods. In the anterior-posterior
differentiation, the accuracy of the AMFM was 89.8% compared to accuracies of 74.6% and 64.4%
by the MLD and HIST methods. In the regional normal vs. emphysema discrimination, over all
the 6 regions, the accuracy of the AMFM was 97.9%. The accuracies using the MLD and HIST
methods were 89.9% and 99.1%, respectively.
The results obtained in this study were substantially better than the results of our previous
study. This demonstrated the importance of standardized scanning protocol.
Further, the Bayesian classifier, being a non-linear
classifier was a better choice for a classifier. In the regional analysis, only those
emphysema regions with definite disease (>80.0%
as assessed visually by an experienced observer)
were used for training and testing.
In recent years, it has been found that emphysema occurs along with interstitial lung disorders
such as idiopathic pulmonary fibrosis (IPF). Therefore, we expanded our study to
characterize IPF in addition to emphysema [3]. MLD was the only method previously
used to describe IPF. The AMFM was compared against the MLD and
HIST methods in simultaneously differentiating normal subjects from subjects with emphysema and IPF.
- Data
20 normal, 10 emphysema, and 19 IPF subjects were included in the study. The normal and IPF
subjects were scanned prone and the emphysema subjects were scanned supine. Data with 3 mm
slice thickness was acquired at full inspiration with an Imatron C-150 XL EBCT scanner.
Four slices from each subject were used for the analysis.
- Methods
The same 17 texture features as in the previous two studies were used. The optimal feature
selection
followed the application of the Bayesian classifier.
- Findings
A global analysis study including whole lungs from CT slices of all the subject groups was
performed. The classification rates of AMFM in identifying
emphysema, IPF, and normal were 100.0%, 97.2%, and 100.0% respectively. The classification
rates of the MLD method for the same three subject groups were 100.0%, 77.8%, and
65.0%. With the HIST method, the classification rates for the same three subject groups
were 80.0%, 61.1%, and 97.5%.
From this study, we concluded that the AMFM was applicable and successful in the detection of
other parenchymal lung diseases in addition to emphysema. Further, the AMFM out-performed
the MLD and HIST methods substantially in simultaneously discriminating normal from emphysema
and IPF.
We progressed [4] towards regional quantitative assessment of the pulmonary
parenchyma. This was accomplished by characterizing the
parenchymal patterns through which the diseases manifest themselves, rather than the
diseases themselves. Sarcoidosis, another interstitial lung disorder, was added to the study.
Representative samples of 6 patterns from normal, emphysema, IPF, and sarcoidosis were identified
on CT by experienced observers. These were honeycombing, ground glass, broncho-vascular, nodular,
emphysema, and normal. Through a training and testing phase, it was ascertained that these
six patterns could successfully be discriminated by the AMFM. Following this, analysis of CT
slices using overlapping
31 x 31
pixel windows was performed. The classification (one
of the six patterns) assigned for each
31 x 31
pixel window was assigned to a
15 x 15 pixel block at its center. The composition of the CT slice in terms of the distribution of
patterns was available using this method. This computer analysis was validated against 5
independent experienced observers.
- Data
12 CT slices derived from normal, emphysema, IPF, and sarcoidosis subjects were included in the study.
The normal, IPF, and sarcoidosis subjects were scanned prone and the emphysema subjects were
scanned supine. Data with 3 mm
slice thickness was acquired at full inspiration with an Imatron C-150 XL EBCT scanner.
- Method
In addition to the 17 features used so far, 5 novel in-house developed stochastic fractal
dimension features were extracted from the data. In total, 22 texture features were used.
The Bayesian classifier was used for the
classification following optimal feature selection.
- Findings
The observers were asked to label the same regions of the lung for which the computer provided
a classification. They performed the labeling in 3 repeat settings, 2-4 weeks apart. The
first and second settings were double-blinded (blinded to primary diagnosis of the subject
and to the computer output). The third setting was single-blinded (primary diagnosis disclosed
but blinded to the computer output). The average inter-observer agreement in the first and second
settings was 43.87.9% and 45.013.8%. The average computer-observer agreement for
the same settings was 36.65.6% and 39.06.0%, respectively. The intra-observer
agreement ranged from 46.2%-77.1%. In the single-blinded study, the average computer-observer
agreement significantly increased (p<0.005) to 48.65.2%. The average inter-observer
agreement was 49.77.7%.
The computer-observer agreement significantly increased in the third setting. This showed
that, while the computer result did not change (being 100% reproducible) the observers
changed their readings based on the knowledge of the primary diagnosis. This suggested that
the AMFM was perhaps correctly identifying the different pathologies. However, the implications
of this method are difficult to appreciate a this point due to lack of gold standards.
Considering that the AMFM performed as well as an observer, it is a valuable tool because
it is reproducible, objective, and quantitative.
So far, all the studies discussed used subjects with well characterized disease.
AMFM proved to be successful in identifying disease or tissue pathologies associated in
such a population. In order to test the sensitivity of the AMFM, we now applied it to a group
of subjects who were smokers but had no pulmonary function abnormalities (SMK)[5].
The goal of this study was to determine if any differences could be detected in this
subject group as compared to two other subject groups: non-smokers with no pulmonary
function abnormalities (NONSMK) and smokers with chronic obstructive pulmonary disorder (COPD).
- Data
Data included 25 CT slices each of NONSMK, SMK, and COPD subjects. Subjects were scanned
in the supine position and 1 mm thick slices at full inspiration were acquired by a Siemens Somatom
DRH scanner.
- Method
The texture feature set was further expanded to include other descriptors of the grey level
histogram. These 5 additional features were based on the percentiles of the histogram. In total,
27 features were used. As before, the Bayesian classifier was used for classification preceded
by the optimal feature selection.
- Findings
In a global analysis study, NONSMK vs. COPD group was differentiated by the AMFM with an
accuracy of 86.0%. The NONSMK vs. SMK group was discriminated with an accuracy of 82.0%
and the SMK vs. COPD groups was discriminated with an accuracy of 70.0%. In a regional
analysis, experienced observers traced samples of the following patterns on CT: normal
lung from NONSMK subjects (norm-NONSMK), normal-appearing lung from COPD subjects (norm-COPD),
and emphysematous lung from COPD subjects (emp-COPD). These three groups were simultaneously
discriminated with an accuracy of 92.0% by the AMFM. The composition of CT slices from
SMK subjects were analyzed using overlapping
31 x 31
pixel windows. It was found that
47.6% of the SMK lung is composed of norm-NONSMK pattern compared to a 50.1% and 2.3%
composed of norm-COPD and emp-COPD patterns, respectively. More recent results
included with images are presented in our companion paper in this SPIE proceedings[8].
This study demonstrated that CT slices of smokers with no lung function abnormalities were
indeed different from CT slices of non-smokers. Also, what appeared ``normal" in COPD
subjects was different from normal lung in non-smokers. We can conclude that the AMFM
is very sensitive to subtle changes which are not detected by the human eye.
A recent study [6] involving asbestos-exposed subjects supported our previous finding that the
AMFM was sensitive to subtle changes. CT images of subjects with significant occupational
exposure to asbestosis were compared to CT images of normal subjects. Some of the
asbestos-exposed subjects had developed asbestosis as confirmed by chest radiograph.
Comparisons were made to see if the AMFM could differentiate normal subjects from asbestos-exposed
subjects.
- Data
CT scans of 20 normal and 46 asbestos-exposed subjects were included. An Imatron C-150 XL
EBCT scanner was used to obtain 3 mm thick slices at full inspiration in prone body posture.
Among the asbestos-exposed subjects, 11 had developed asbestosis. Four slices from each
subject were used in the analysis.
- Method
The texture features used were the same as the 27 used in the previous study. The classification
was performed by a Bayesian classifier after the optimal feature selection.
- Findings
The sensitivity and specificity in the asbestos-exposed vs. normal discrimination task were
93.5% and 87.5%, respectively. In differentiating asbestosis vs. normal, the sensitivity and
specificity were 90.9% and 97.9%. Finally, in discriminating asbestos-exposed subjects who
appeared normal on chest radiographs vs. normal subjects, the sensitivity and specificity were
90.0% and 87.5%.
This study provided an important result. The AMFM was successful in differentiating subjects
with normal chest radiographs who had and who had not been exposed to asbestos. Current screening
techniques for occupational exposure to asbestosis only include routine chest radiographs.
The AMFM can serve as a useful tool in screening CT scans of subjects who have minimal
radiographic evidence of asbestosis.
It is critical and urgent that better methods be developed to identify
and measure early lung injury in cystic fibrosis (CF) as it has become very clear that the
end stage lung pathology in CF patients is largely due to early repeated
infectious insults to the lung. The plain chest radiograph and pulmonary
function testing are currently the standard measures of lung injury,
but, neither is particularly sensitive to early changes occuring prior
to the time that bronchiectasis (distal airway dilation) becomes well
established. Bronchoalveolar lavage (BAL) data serve as a markers of
early infection and inflammation.
- Data
Three CF
patients who had minimal to no clinical signs of lung disease but had
BAL based measures showing: Subj. 1) neutrophil count (NC) of 2 and no
growth (NG); Subj. 2) NC of 15 and NG; and Subj. 3) a NC of 43 with
pseudomonas and staph respectively. Subjects were scanned via EBCT with
a 3 mm slice thickness. A high resolution reconstruction algorithm was
applied and 8-12 slices were gathered, evenly spaced from lung
apex-to-base.
- Method
While the AMFM method most certainly
needs to be trained using an appropriate age matched population of
children with normal lung, our AMFM approach, trained previously
to recognize the six tissue types associated with emphysema, IPF, sarcoidosis, and normal lung,
was preliminarily applied in a retrospective analysis.
- Findings
Application of the AMFM method to the CT slices gathered
from these subjects labeled 77.03.0%, 33.05.0%, and 8.03.0% of the per slice
lung fields as normal for subjects 1-3, respectively. Furthermore, the
heterogeneity of regional tissue labels was maximal for subject 2 and
heterogeneously ground glass for subject 3. An assessment of the
relationship between airway luminal diameter (LD) vs wall thickness (WT)
was assessed for airways ranging in LD from 2 to 12mm. The slope (WT/LD)
of the relationship for subjects 1 and 2 was 0.08 and 0.09 while the slope
for subject 3 was 0.32, suggesting early bronchial wall thickening and
bronchiectasis in subject 3.
We have not drawn any statistical
conclusions from these data sets. However, when they are viewed in light
of our successes in other lung diseases and our significant advances in
the development of software for the evaluation of airway morphology and
physiology data, these results suggest that this approach shows
significant promise for serving as a quantitative approach to the
detection and quantification of early lung changes in CF [7].
The AMFM is the only available comprehensive lung assessment tool
for CT. It is automated, objective, and completely reproducible. It has been tried and tested
in a variety of scenarios and has proved to be successful as a reliable tissue characterization
technique.
The first design of the method included texture features, most of which were available
and had been applied in other pattern recognition problems before. The geometric fractal dimension
was developed specifically for the lung. As the method evolved, more new features were added.
The stochastic fractal dimension measures, based on the fractional Brownian motion model
were also used. These measures substantially added to the discriminatory power of the AMFM.
Finally based on the previous success of the lowest fifth percentile of the histogram (HIST)
as a measure of emphysema, 5 percentile measures were developed. The classifier was also
changed early on from a minimum distance classifier to a Bayesian classifier. The Bayesian
classifier is a non-linear classifier based on probability density distribution model
and minimum loss optimality criterion.
The analysis started with global analysis utilizing whole lung slices to detect presence
or absence of disease. In discrimination tasks with multiple subjects groups, the AMFM
identified slices with each disease simultaneously. The global analysis was later followed
by regional analysis which indicated which regions of the CT slice were diseased. This
also allowed for the analysis of the composition of patterns in CT slices.
The data used for the first study did not use any standardized scanning protocols. However,
in all following studies, scanning was standardized. Most studies used 3 mm thick slices.
One study used 1 mm. The AMFM was successful using both high-resolution slices. Imatron
and Siemens scanners were used in different studies. This demonstrated that the AMFM is
applicable across scanners. Some studies used subjects with well characterized disease
while others used subjects with no apparent disease. The AMFM was successful in differentiating
all these subjects from normal controls.
Future work includes performing volumetric scans and analyzing the lung using the AMFM
in 3 dimensions. Efforts are also on to reduce the analysis time required by the computer.
We can conclude based on the above studies that the AMFM is a tool that can be
used for screening early or well characterized disease. It is automated, objective,
and quantitative and hence can be used in clinical as well as research settings. Such
a method can be used to assess changes in longitudinal studies and can allow for
quantitative cross-institutional outcomes measures.
Supported in part by the a contract from the National Library
of Medicine (N01-LM-4-3511 US PHS) and a career investigator award from the
American Lung Association. Reprint requests to: Eric A. Hoffman, PhD, Department
of Radiology, University of Iowa College of Medicine, 200 Hawkins Dr., Iowa City, IA, 52242.
- 1
-
R. Uppaluri, T. Mitsa, E. A. Hoffman, G. McLennan, and M. Sonka, ``Texture
analysis of pulmonary parenchyma in normal and emphysematous lung,'' in Medical Imaging 1996: Physiology and Function from Multidimensional Images,
Vol. 2709, Newport Beach, CA, pp. 456-467, SPIE, (Bellingham, WA), 1996.
- 2
-
R. Uppaluri, T. Mitsa, M. Sonka, E. A. Hoffman, and G. McLennan,
``Quantification of pulmonary emphysema from lung computed tomography
images,'' Am. J. Respir. Crit. Care Med. 156(1), pp. 248-254,
1997.
- 3
-
R. Uppaluri, M. Sonka, E. A. Hoffman, T. Mitsa, C. Dayton, G. W. Hunninghake,
and G. McLennan, ``Quantitative assessment in interstitial lung disease using
the multiple feature (MF) method (abstract),'' Am. J. Respir. Crit.
Care Med. 155(4), p. A322, 1997.
- 4
-
R. Uppaluri, P. G. Hartley, E. A. Hoffman, M. Sonka, and G. McLennan,
``Validation of AMFM: A quantitative, regional assessment of pulmonary
parenchymal pathology from HRCT (abstract),'' Radiology 205((P)), p. 481, 1997.
- 5
-
R. Uppaluri, G. McLennan, and E. A. Hoffman, ``AMFM - a quantitative CT
assessment of early parenchymal changes in smokers (abstract),'' Am. J.
Respir. Crit. Care Med. (in press) , 1998.
- 6
-
R. Uppaluri, E. A. Hoffman, D. A. Schwartz, G. McLennan, G. W. Hunninghake,
C. Dayton, and P. G. Hartley, ``Quantitative analysis of the chest CT in
asbestos-exposed subjects using the Adaptive Multiple Feature Method
(abstract),'' Am. J. Respir. Crit. Care Med. (in press) , 1998.
- 7
-
E. A. Hoffman and G. McLennan, ``Needs for assessment of pulmonary
structure/function relationships and clinical outcomes measures:
Quantitative volumetric CT of the lung,'' Academic Radiology 4(11), pp. 758-776, 1997.
- 8
-
R. Uppaluri, G. McLennan, P. Enright, J. R. Standen, P. Boyer-Pfersdorf, and
E. A. Hoffman, ``Adaptive Multiple Feature Method (AMFM) for the
early detection of parenchymal pathology in a smoking population,'' in Medical Imaging 1998: Physiology and Function from Multidimensional Images,
Vol. 3337, San Diego, CA. (in press), SPIE, (Bellingham, WA), 1998.
©1994-2000 Division of Physiologic Imaging, Dept.
of Radiology, Univ. of Iowa
DPI Homepage |
VIDA |
NLM |
Contact Us |
Search
Last modified: Wed May 31 15:29:49 CDT