Adaptive Multiple Feature Method (AMFM) for the early detection of parenchymal pathology in a smoking population

Renuka Uppaluri1, Geoffrey McLennan2, Paul Enright3,
James R. Standen4, Pamela Boyer-Pfersdorf3, and Eric A. Hoffman1,5

1Department of Radiology University of Iowa, Iowa City, IA 52242
2Department of Internal Medicine University of Iowa, Iowa City, IA 52242
3Department of Internal Medicine University of Arizona, Tucson, AZ 85724
4Department of Radiology University of Arizona, Tucson, AZ 85724
5Department of Biomedical Engineering University of Iowa, Iowa City, IA 52242


[ home | search | contact us ]


 

Table of Contents

Abstract:

Application of the Adaptive Multiple Feature Method (AMFM) to identify early changes in a smoking population is discussed. This method was specifically applied to determine if differences in CT images of smokers (with normal lung function) and non-smokers (with normal lung function) could be found through computerized texture analysis. Results demonstrated that these groups could be differentiated with over 80.0% accuracy. Further, differences on CT images between normal appearing lung from non-smokers (with normal lung function) and normal appearing lung from smokers (with abnormal lung function) were also investigated. These groups were differentiated with over 89.5% accuracy. In analyzing the whole lung region by region, the AMFM characterized 38.6% of a smoker lung (with normal lung function) as mild emphysema. We can conclude that the AMFM detects parenchymal patterns in the lungs of smokers which are different from normal patterns occurring in healthy non-smokers. These patterns could perhaps indicate early smoking-related changes.

Tissue characterization, texture, smoking, pulmonary, Bayesian classifier

INTRODUCTION

We have made significant advances in the use of tissue characterization to classify parenchymal patterns in reconstructed CT data sets. The Adaptive Multiple Feature Method (AMFM) is a texture classification scheme that has been successfully applied towards discrimination of various tissue types associated with pulmonary emphysema, idiopathic pulmonary fibrosis, and sarcoidosis [1,2,3]. While all our previous studies involved CT images of subjects with well characterized disease, this study seeks to evaluate early parenchymal pathology associated with smoking. This is a challenge because early changes are subtle making them difficult to detect.

Smoking is one of the main causes of debilitating diseases such as emphysema. Emphysema is a form of chronic obstructive pulmonary disorder (COPD). The onset of emphysema starts early but it presents itself clinically at a stage when it is incurable and care is largely supportive. In recent years, lung volume reduction surgery has become a viable intervention for end-stage emphysema subjects. However, patient selection criteria are poorly defined and, more importantly, sensitive measures for outcome evaluation are lacking. Novel anti-elastase and other anti-inflammatory modulators are being developed to halt the progression of early emphysema. Therefore, the motivation for this study is to identify a means of detecting changes caused by smoking before clinical symptoms appear. This can then be used either to advise the patient to adopt lifestyle changes such as smoking cessation or to determine the outcome of therapeutic interventions. It should be noted, once established, emphysema may progress even with smoking cessation.

In this study, three groups of subjects are included - healthy non-smokers with normal lung function, healthy smokers with normal lung function, and smokers with abnormal lung function or COPD. The AMFM will be used to demonstrate the differences between these three groups.

METHODS

The AMFM is described in detail elsewhere [1]. To summarize, the AMFM is a pattern recognition algorithm which performs recognition based on a learning (training) stage. The different patterns are described via a set of texture measures. During the training stage, a) an optimal subset of features is determined, which best discriminates the patterns in training set and b) a classification algorithm is trained to perform the discrimination of the patterns in the training set using the optimal subset of features. During the testing stage, the accuracy of the AMFM is tested on an independent set of patterns.

The optimal feature selection is performed by using the divergence measure along with correlation analysis [4] and the classification is performed by a Bayesian approach. The texture measures used are derived from various classes. The first order measures based on the grey level histogram, run-length measures, the co-occurrence matrix measures, and the geometric fractal dimension computations are presented in [5,6,1]. In addition to these 17, 10 new measures, developed by us, are included in this study. They are the percentile measures, also based on the grey level histogram, and the stochastic fractal dimension measures [7]. All measures used and their abbreviations are provided in Table 1.


  
Table 1: Texture measures and their abbreviations
Feature Abbreviation
Mean MEAN
Variance VAR
Skewness SKEW
Kurtosis KURT
Grey level entropy GREYENT
Short run emphasis SRE
Long run emphasis LRE
Grey level non-uniformity GLN
Run length non-uniformity RLN
Run Percentage RP
Angular second Moment ASM
Inertia INER
Contrast CON
Entropy ENT
Correlation CORR
Inverse difference moment IDM
Lowest fifth percentile LOW-FIFTH
Upper fifth percentile UPPER-FIFTH
Mean - Lowest fifth percentile LOWER-DIFF
Upper fifth percentile - mean UPPER-DIFF
Ratio of UPPER-DIFF to LOWER-DIFF DIFF-RATIO
Geometric fractal dimension GFD
Stochastic fractal dimension mean SFD MEAN
Stochastic fractal dimension variance SFD VAR
Stochastic fractal dimension skewness SFD SKEW
Stochastic fractal dimension kurtosis SFD KURT
Stochastic fractal dimension grey level entropy SFD ENT

EXPERIMENTAL METHODS

Subjects and Image Acquisition

The subjects included in this study were: 41 healthy non-smokers (NONSMK), 24 smokers (SMK), and 19 smokers with abnormal lung function (COPD).

A Siemens DRH Somatom scanner was used to acquire CT scans of all subjects in supine position. Two slices (1 mm slice thickness) were acquired at full inspiration 5 cm above and 5 cm below the carina. The images were reconstructed to 512$\times$512 pixels and the grey levels were rescaled to lie between 0-2048.

Measurements and Validation

1.
Global analysis:

All three subject groups were used in this experiment. The left and right lungs from each image were isolated and used individually as the region of interest for texture computation and classification. Half the lungs were put into a training set and the remaining were put into a test set. The goal was to discriminate a NONSMK lung from SMK lung, a NONSMK lung from a COPD lung, and a SMK lung from a COPD lung. The ability to discriminate simultaneously between a NONSMK lung, SMK lung, and COPD lung was also tested.

On the test set, classification accuracy of each subject group and the overall accuracy of all subject groups was computed for each discrimination task. The classification accuracy is the percent of samples belonging to a group correctly classified as such. The overall accuracy is the percent of correctly classified samples of all groups together.

2.
Regional analysis (observer-defined regions):

In this experiment, NONSMK and COPD subject groups were used. A trained observer identified areas of four tissue types, namely, normal lung from NONSMK subjects (NN), emphysema-like lung from COPD subjects (EC), ``normal" lung from COPD subjects (NC), and mild emphysema-like lung from COPD subjects (MC). In each defined region, pixel blocks of 31$\times$31 overlapping by 15$\times$15 were placed. Half the blocks were put into a training set and the remaining were put into a test set. The goal was to discriminate NN samples from each of EC, NC, and MC samples, individually. The ability to discriminate between NN from EC, NC, MC simultaneously was also tested.

As before, classification accuracies and overall accuracies were computed.

3.
Regional analysis (whole lung slices):

All three subject groups were included in this experiment. On each CT slice, 31$\times$31 pixel blocks overlapping by 15$\times$15 were translated through the lung regions and the pattern classification obtained for each block was assigned to the 15$\times$15 block at its center. The classifier assigned either a NN, EC, NC, or MC pattern using the training acquired in the previous experiment. Each block was color-coded based on its pattern and the images were displayed. The goal of this experiment was to determine what percent of a NONSMK, SMK, and COPD lung is comprised of each of the four patterns.

RESULTS

Global analysis

The results for this experiment are partially shown in Table 2. The training and test sets contained 25 samples of NONSMK, SMK, and COPD groups. In discriminating all three groups simultaneously, the classification accuracies were 80.0%, 52.0%, and 64.0% for the NONSMK, SMK, and COPD groups, respectively. The overall accuracy here was 65.3%.


  
Table 2: Global analysis results. NONSMK: Healthy non-smokers, SMK: Smokers with normal lung function, COPD: Smokers with COPD.
Discrimination Classification Classification Overall
task Accuracy (Group 1) Accuracy (Group 2) Accuracy
Group 1: NONSMK 88.0% 84.0% 86.0%
vs. Group 2: COPD      
Group 1: NONSMK 88.0% 72.0% 80.0%
vs. Group 2: SMK      
Group 1: SMK 76.0% 72.0% 74.0%
vs. Group 2: COPD      

Regional analysis (observer-defined regions)

The results for this experiment are partially shown in Table 3. There were 180 NN, 180 EC, 48 NE, and 76 MC samples in the training and test sets. In discriminating all four groups together, the classification accuracies were 86.7%, 92.2%, 68.7%, and 77.6%, for the NN, EC, NC, and MC groups, respectively. The overall accuracy here was 85.5%.


   
Table 3: Regional analysis results (observer-defined regions). NN: Normal lung region from healthy non-smokers, EC: Emphysematous lung region from smoking COPD subjects, NC: Normal appearing lung region from smoking COPD subjects, MC: Mild emphysematous lung region from smoking COPD subjects.
Discrimination Classification Classification Overall
task Accuracy (Group 1) Accuracy (Group 2) Accuracy
Group 1: NN 99.4% 97.8% 98.6%
vs. Group 2: EC      
Group 1: NN 94.4% 70.8% 89.5%
vs. Group 2: NC      
Group 1: NN 96.7% 88.1% 94.1%
vs. Group 2: MC      

Regional analysis (whole lung slices)

The results of this experiment are shown in Table 4. The table shows the distributions of NN, EC, NC, and MC patterns in NONSMK, SMK, and COPD lungs. Examples of a NONSMK, SMK, and COPD lung with their computer outputs are shown in Figure 1.


  
Figure 1: Examples of lungs with their color-coded computer output. (a) An healthy non-smoker's lung (NONSMK) (b) Computer output of NONSMK lung (c) An smoker's lung who has normal lung function (SMK) (d) Computer output of SMK lung (e) Lung of a smoker with abnormal lung function (COPD) (f) Computer output of the COPD lung. NN: Normal lung region from healthy non-smokers, EC: Emphysematous lung region from smoking COPD subjects, NC: Normal appearing lung region from smoking COPD subjects, MC: Mild emphysematous lung region from smoking COPD subjects.
\begin{figure}\centerline{ $ \begin{array}{cc} \psfig{figure=Figs/NONSMK.ps,heig... .../COPD.out.ps,height=1.5in,width=1.5in}\\ (e)&(f)\\ \end{array}$ } \end{figure}


   
Table 4: Regional analysis results (whole lung slices). NONSMK: Healthy non-smokers, SMK: Smokers with normal lung function, COPD: Smokers with COPD, NN: Normal lung region from healthy non-smokers, EC: Emphysematous lung region from smoking COPD subjects, NC: Normal appearing lung region from smoking COPD subjects, MC: Mild emphysematous lung region from smoking COPD subjects.
Subject NN EC NC MC
group        
NONSMK 39.7% 13.9% 16.5% 29.9%
SMK 22.3% 15.2% 23.9% 38.6%
COPD 9.5% 40.3% 4.8% 35.4%

DISCUSSION

There are several issues that need to be addressed in this study. The first one is the lack of gold standards. Emphysema can be characterized by bullae and mild emphysema can be characterized by subtle changes in the lung architecture, visible to the eye on CT. So there are visual and clinical standards for these. However, there are no standards to detect pathologies related to smoking. Therefore, our goal was to determine if such pathologies, that we postulate to occur in the lung as a result of smoking, are detectable by computerized texture analysis via CT.

The AMFM has demonstrated to be successful in differentiating normal subjects from subjects with severe emphysema [1], and more recently with interstitial lung diseases such as idiopathic pulmonary fibrosis, and sarcoidosis [2]. Since, in this study, NONSMK (non-smokers with normal lung function) can be differentiated from SMK (smokers with normal lung function) with 80.0% accuracy, this further demonstrates that the AMFM is sensitive enough to pick up subtle differences. Our regional analysis experiments were used to localize and characterize the differences between the groups. It was first demonstrated that visually normal appearing lung regions from NONSMK and COPD subjects (smokers with abnormal lung function) were differentiated with 89.5% accuracy. An explanation for this could be that normal appearing lung in COPD subjects appears ``normal" to the observer only because it is juxtaposed with severely diseased lung. The mild emphysematous regions and emphysema regions were also discriminated from normal lung of NONSMK with accuracies higher than 94.0%.

In analyzing whole lung slices, it was found that the majority of a SMK lung was composed of mild emphysema and ``normal from COPD", i.e., lung which visually appears normal but is different from normal lung of non-smokers. It is also noted that 22.9% of a NONSMK lung is composed of mild emphysema. We attribute this to the presence of dark regions juxtaposed to ribs which may be occurring due to beam hardening. Since such areas were not included in the training set as being normal, the computer assigned labels of mild emphysema or emphysema. The computer can perform only as well as the training it receives. Therefore, representative training examples along with good data acquisition using standardized protocols are of utmost importance. Futhermore, it is critical that motion be eliminated during scanning and that lung volume be standardized.

This study also demonstrates the applicability of the AMFM across scanners. While our previous studies were performed on Imatron C-150 scanner with 3 mm thick slices, this study uses 1 mm thick slices from Siemens Somatom DRH scanner. It should be noted that, for best results, the AMFM should be re-trained for each scanner and for each new set of scanning and reconstruction parameter.

CONCLUSION

This study shows that a) via CT, the smokers with normal lung function can be differentiated from healthy non-smokers and smokers with COPD b) normal appearing lung regions from COPD subjects can be differentiated from normal lung regions from healthy non-smokers c) in smokers with normal lung function, there is a higher occurrence of the mild emphysema pattern. We conclude that the lung from smokers with normal lung function is indeed abnormal, suggesting that there are changes which can be attributed to smoking. It is our hypothesis that these changes pre-date emphysema. Longitudinal studies would be required to answer this, and the AMFM can be used for such a study.

ACKNOWLEDGEMENTS

Supported in part by the a contract from the National Library of Medicine (N01-LM-4-3511 US PHS), NIH NHLBI SCOR (HL-14136), and a fellowship grant from the Cystic Fibrosis Foundation. Reprint requests to: Eric A. Hoffman, PhD, Department of Radiology, University of Iowa College of Medicine, 200 Hawkins Dr., Iowa City, IA, 52242.

References

1
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.

2
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.

3
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.

4
H. C. Andrews, Introduction to mathematical techniques in pattern recognition, Wiley, New York, 1972.

5
R. Uppaluri, T. Mitsa, and J. R. Galvin, ``Fractal analysis of high-resolution CT images as a tool for quantification of lung diseases,'' in Medical Imaging 1995: Physiology and Function from Multidimensional Images, Vol. 2433, San Diego, CA, pp. 133-142, SPIE, (Bellingham, WA), 1995.

6
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.

7
R. Uppaluri, Automated analysis of pulmonary parenchyma from CT images.
PhD thesis, University of Iowa, Iowa City, Iowa, USA, 1997.




DPI Homepage | VIDA | NLM | Contact Us | Search


Last modified: Thu Apr 16 17:03:49 CDT