CARS 2002 H.U. Lemke, M.W. Vannier; K. Inamura, A.G. Farman, K. Doi & J.H.C. Reiber (Editors)

CARS 2002 – H.U. Lemke, M.W. Vannier; K. Inamura, A.G. Farman, K. Doi & J.H.C. Reiber (Editors)

CARS/Springer. All rights reserved.

1

Automatic segmentation and texture analysis of PA chest radiographs to detect abnormalities related to interstitial disease and tuberculosis

Bram van Ginnekena, Bart M. ter Haar Romenyb, Max A. Viergevera

a Image Sciences Institute, University Medical Center Utrecht,

Utrecht, The Netherlands

bFaculty of Biomedical Engineering, Medical and Biomedical Imaging, Eindhoven University of Technology, Eindhoven, The Netherlands

Abstract

We present an automatic system for detecting diffuse abnormalities in chest radiographs. The system starts with segmentation, subdivides the lung fields in smaller, overlapping regions and extract texture features from each ROI. Using these features, the probability that each ROI contains abnormalities is estimated with a k-nearest-neighbour classifier. The classification of all regions is pooled into an overall abnormality indicator. Evaluation on databases containing cases of interstitial disease and tuberculosis shows promising results. Directions for further research are briefly discussed.

Keywords: Computer-aided diagnosis, chest radiography, texture analysis.

1.Introduction

Around thirty percent of all radiological examinations are conventional chest radiographs. Given this huge number of studies, and the fact that they may contain extremely subtle abnormalities, it is not surprising that computer-aided diagnosis (CAD) in chest radiography is an active research area with serious potential for clinical applications [1]. We are developing a system for CAD in chest radiographs. It currently focuses on the detection of interstitial abnormalities, and could be used in tuberculosis screening.

Broadly speaking, abnormal signs in chest images can be subdivided in two categories: signs independent of location, and those signs whose appearance depends on their location in the lung field. Lung nodules are an example of the former, and most existing algorithms for lung nodule detection perform an analysis independent of the location of a candidate lesion. Diffuse abnormalities that characterize interstitial disease are examples of the latter. Consequently, algorithms to detect such signs can benefit from a specific analysis for each lung region.

2.Methodology

2.1 Algorithm outline

In order to perform a regional analysis of chest images, we propose the following multi-stage approach. The first step is an automatic segmentation of both lung fields. Subsequently, the lung fields are divided into smaller regions of interest (ROIs). Each region is searched for abnormal signs. Because we focus on diffuse abnormalities, we perform texture analysis and extract a set of features from each ROI. In an off-line stage, the results for each region on images with known locations of abnormalities are used to construct a training set of reference cases. In the on-line analysis, a statistical classifier estimates the likelihood that a specific ROI is abnormal. The system could stop here and present the possibly abnormal regions to a radiologist. Optionally, the results of all regions can be pooled in a final stage into a single score for the complete image.

2.2 Segmentation

Currently we employ a modified version [2] of Active Shape Models [3] to extract the lung fields from a chest radiograph. This method is very robust and trainable, making it easy to adjust it to other databases with different image characteristics. The segmentation result is generally sufficiently accurate for our purposes.

2.3 Subdividing the lung fields

Each lung fields is divided into an upper, middle and lower part (6 regions). Each region is subdivided into a medial and lateral part (12 region), each of which is finally divided into an upper and lower part again (24 regions). Thus the total number of - overlapping - regions is 42. The exact subdivision used is not critical. The notion of overlapping regions is important though, it may allow detection of both small and large abnormal regions. Alternatively, a subdivision into costal and intercostal space could be beneficial, but would require an accurate algorithm for delineating the rib cage.

2.4 Texture feature extraction

A filter-bank texture analysis method that extracts multi-scale texture features from local histograms is used (see e.g. [4]). The set of filters is given by the Gaussian and its derivatives up to second order, at multiple scales. The histogram per region of the filtered image is computed, from which the first four moments are extracted. Density features are added and the difference between corresponding regions in the left and right lung field is used to construct additional ‘difference’ features, in order to mimic right-left comparisons as they are made routinely by radiologists.

2.5 Region classification

The features are each scaled to zero mean and unit variance. A simple k-nearest neighbour classifier is used to estimate the probability that a region is abnormal. Different sets of features (different scales, different moments, whether or not to include the difference features, etc.) have been tested on several databases. A selection of features for each region separately could be employed, but the size of our databases is currently rather small (for some regions the number of features is much larger that the number of abnormal cases), therefore we currently do not perform feature selection.

2.6 Region pooling

For each region, the performance of the system can be measured in terms of Az, the area under the ROC curve. Using Az, we perform a weighted average of all regions to arrive at a final abnormality score of the complete image. The exact way in which the weighting is implemented is ad hoc, but turns out to have little effect on the overall performance of the system. The weighting procedure ensures that regions for which no reliable estimate can be made have only a small influence on the total abnormality estimate.

3.Results

We show results of the system on two databases. The first database, referred to as the TB database, contains 279 abnormal and 290 normal posterior-anterior (PA) chest radiographs collected from a tuberculosis screening program for people seeking political asylum in The Netherlands [5]. The second database, the ID database, contains 100 normal and 100 abnormal PA chest radiographs with interstitial disease obtained at the University of Chicago Hospitals [6].

Fig. 1: ROC curves for both databases. The thin lines below and above the curve denote the asymmetric 95% confidence intervals. (Left) The TB database. The area under the curve is 0.820. (Right) The ID database. The area under the curve is 0.986.

Figure 1 shows ROC curves. For the ID database, the results are near perfect. The TB database contained many subtle abnormalities, which are not easily detected by the system. Furthermore, the abnormal areas were often small (in the order of a few percent of the area of a single lung field) and since the system averages the results over all regions, it is not very sensitive to small abnormalities.

4. Discussion

The use of the system has not yet been evaluated in clinical practice. There are several ways in which such a system could be used in practice: (a) As a stand-alone filtering stage in which each image in a mass chest screening is processed and only sent to a radiologist if it is possibly abnormal; (b) As a computer-aided diagnosis module that the radiologist can use as a second opinion; (c) To highlight possibly abnormal areas during reading; (d) To retrieve images with similar textural appearance in abnormal regions as reference cases. This enumeration is by no means exhaustive.

To make the system more powerful, it should be trained with substantially larger database. This may allow it to catch more subtle abnormal sign that occur over small areas only. A range of improvements and extensions could be integrated into the system. Current research areas are (a) Integration of previous chest radiographs using temporal subtraction techniques; (b) Integrating clinical information (patient history) into the classification stage; (c) Include a dedicated algorithm for the detection of lung nodules and dense infiltrates, for which texture analysis is not the most suitable approach; (d) Detection of the rib cage, the heart and the clavicles prior to analysis; (e) Including shape analysis of segmented objects in the images – abnormalities such as blunting of the costophrenic angle can be handled in this way .

References

[1] B. van Ginneken, B.M. ter Haar Romeny, M.A. Viergever, “Computer-aided diagnosis in chest radiography: a survey”, IEEE Trans. on Medical Imaging, 20(12):1228-1241, 2001.

[2] B. van Ginneken, A.F. Frangi, J.J. Staal, B.M. ter Haar Romeny, M.A. Viergever, “A non-linear gray-level appearance model improves active shape model segmentation”, in IEEE Workshop on Mathematical Models in Biomedical Image Analysis (MMBIA 2001), pp. 205-212.

[3] T.F. Cootes, C.J. Taylor, D. Cooper, J. Graham, “Active shape models - their training and application”, Computer Vision and Image Understanding, 61(1):38-59, 1995.

[4] M. Unser. and M. Eden, “Multi-resolution feature extraction and selection for texture segmentation”, IEEE Trans. on Pattern Analysis and Machine Intelligence, 11(7):717-728, 1989.

[5] B. van Ginneken, S. Katsuragawa, B.M. ter Haar Romeny, K. Doi, M.A. Viergever, “Automatic detection of abnormalaties in chest radiographs using local texture analysis”, IEEE Trans. on Medical Imaging, 21(2), to appear, 2002.

[6] T. Ishida, S. Katsuragawa, K. Ashizawa, H. MacMahon, K.Doi, “Application of artificial neural networks for quantitative analysis of image data in chest radiographs for detection of interstitial lung disease”, Journal of Digital Imaging, 11(4):182-192, 1998.