Download Automated Detection of Blunt Costophrenic Angles in Chest Radiographs and more Study Guides, Projects, Research Radiology in PDF only on Docsity!
Automated localization of costophrenic recesses and
costophrenic angle measurement on frontal chest radiographs
Pragnya Maduskar, Laurens Hogeweg, Rick Philipsen,
and Bram van Ginneken
Diagnostic Image Analysis Group, Department of Radiology,
Radboud University Nijmegen Medical Centre, The Netherlands
ABSTRACT
Computer aided detection (CAD) of tuberculosis (TB) on chest radiographs (CXR) is difficult because the disease has varied manifestations, like opacification, hilar elevation, and pleural effusions. We have developed a CAD research prototype for TB (CAD4TB v1.08, Diagnostic Image Analysis Group, Nijmegen, The Netherlands) which is trained to detect textural abnormalities inside unobscured lung fields. If the only abnormality visible on a CXR would be a blunt costophrenic angle, caused by pleural fluid in the costophrenic recess, this is likely to be missed by texture analysis in the lung fields. The goal of this work is therefore to detect the presence of blunt costophrenic (CP) angles caused by pleural effusion on chest radiographs.
The CP angle is the angle formed by the hemidiaphragm and the chest wall. We define the intersection point of both as the CP angle point. We first detect the CP angle point automatically from a lung field segmentation by finding the foreground pixel of each lung with maximum y location. Patches are extracted around the CP angle point and boundary tracing is performed to detect 10 consecutive pixels along the hemidiaphragm and the chest wall and derive the CP angle from these.
We evaluate the method on a data set of 250 normal CXRs, 200 CXRs with only one or two blunt CP angles and 200 CXRs with one or two blunt CP angles but also other abnormalities. For these three groups, the CP angle location and angle measurements were accurate in 91%, 88%, and 92% of all the cases, respectively. The average CP angles for the three groups are indeed different with 71.6◦^ ± 22.9, 87.5◦^ ± 25.7, and 87.7◦^ ± 25.3, respectively.
Keywords: CAD, Chest X-ray, Costophrenic angle, Costophrenic recess, Pleural effusion, Pleural fluid, Radio- graph, Tuberculosis
1. PURPOSE
Detection of tuberculosis (TB) in chest radiographs (CXRs) is challenging as the abnormalities that can be visible on the image (opacifications, pleural effusions, blunt costophrenic angle, lymphadenopathy, hilar elevation) span a wide range. A CAD prototype for detection of TB (CAD4TB v1.08, Diagnostic Image Analysis Group, Nijmegen, The Netherlands)^1 was developed which is trained to detect parenchymal abnormalities inside the unobscured lung fields. Abnormalities like pleural effusion, blunt costophrenic angle and hilar elevation can be missed by this system. Various CXR scoring systems for TB have been developed in which the user provides a score for a number of specific abnormalities that may or may not be present in the image.2, 3^ Existence of pleural effusion or blunt costophrenic angle is one of the important abnormalities consistent with TB.4, 5^ If this is the only abnormality present in the image, it might go undetected in computerized analysis of the chest radiograph. To detect such abnormalities, a promising strategy is to develop sub-systems which are specifically trained to detect a particular type of abnormality. Outputs of these sub-systems can then be combined into a CAD system to detect all the abnormalities suggestive of TB. This work is aimed towards detecting specifically blunt costophrenic angles.
The costophrenic (CP) angle is formed by the lateral chest wall and the dome of each hemidiaphragm. Blunt CP angle is often caused by pleural effusion, where the costophrenic recesses are filled with pleural fluid. Sometimes,
Further author information: (Send correspondence to P.Maduskar@rad.umcn.nl)
pleural fluid is not apparent and only a blunt CP angle is visible. Not much research has been done on automatic detection of blunt CP angle so far. There is only one proposed technique^6 in literature to automatically delineate costophrenic angles. It uses iterative gray-level thresholding for lung segmentation as an initial step. CP angle measurements were performed after delineating the diaphragmatic and costal aspect of the CP angle using contrast information. It was tested on 1,166 hemithoraces where the task was to classify the image as containing a blunt CP or not, and area under the receiving operating curve of 0.83 was reported for severely blunt costophrenic angle.
In our technique, we use a supervised learning approach for detecting lung fields automatically. Accuracy of localization of left and right costophrenic recesses is reported. Angle measurements are performed for both hemithoraces to identify the range of CP angle values for normal images and abnormal images with blunt CP angle. This work is the first step towards a sub-system for automatic detection of blunt costophrenic angle and pleural effusion in a CXR image. This sub-system can be then combined with the CAD system^1 for detection of TB.
2. METHODS
Our approach consists of three steps: Automatic lung fields segmentation; costophrenic recess localization; costophrenic angle measurement. They are detailed below.
2.1 Automatic lung fields segmentation
A pixel classification approach similar to a previously published method^7 is used for automatic lung segmentation. We used a training data set of 299 digital CXRs (Delft Imaging Systems, Veenendaal, The Netherlands) containing both normal and abnormal images to accurately segment lungs in abnormal images as well. Lung fields were manually annotated by a training reader and pixels were randomly sampled inside and outside the lung fields to train the pixel classifier. To capture the texture of the lung parenchyma, the output of Gaussian derivative filtered images of order 0 through 2 (L, Lx, Ly , Lxx, Lxy , Lyy ), at scales 1, 2, 4, 8 and 16 pixels were used as features. Absolute x and y location of pixels was also used as features as position information. Normalized features were used to label the pixels as lung or background using k-nearest neighbor classifier (k=15). All the calculations for lung fields segmentation are performed on images scaled to a width of 256 pixels.
2.2 Costophrenic recess localization
We detect the CP angle point which is defined as the intersection point of the hemidiaphragm and the lateral chest wall. CP angle point is located by scanning the lung mask (left and right separately) in the x direction from the bottom of the image moving upwards until a foreground pixel (im(x, y) = 1) is found. A few examples of detected CP angle point are shown in Fig 1. This operation is performed for left and right lung segmentation masks separately to detect the corresponding left and right CP angle point. We extract a patch of 100× 100 pixels around the CP angle point such that two-third of the pixels are above the CP angle point. Left and right patches around the CP angle point are shown in Fig. 4.
2.3 Costophrenic angle measurement
Costophrenic angle measurement is performed on these patches by tracing the contours along the lung segmen- tations. These correspond to the hemidiaphragm and the lateral chest wall. We start with the CP angle point and scan in x and y directions to mark the first foreground pixel hit on the lung mask. This is done for 10 con- secutive rows and columns starting at the CP angle point, which gives 10 consecutive points on hemidiaphragm and lateral chest wall. We average these pixel locations to obtain a single point on both the contours (Fig. 2). The measurements are performed on 1024 pixel width resolution image with isotropic pixel size of 0.45 mm. Finally the costophrenic angle is calculated between the two lines defined by a line between CP angle point - hemidiaphragm point and CP angle point - lateral chest wall point (Fig. 2). Angle measurement is done using Eq. 1 where m 1 and m 2 are slopes of the two lines.
θ = | arctan(m 1 ) − arctan(m 2 )| (1)
40
60
80
100
120
140
160
Normals
Blunt CP+Abnormal
Costophrenic angle (in degrees)
Blunt CP only
Figure 3. Box plot illustrating range of costophrenic angles for normal and abnormal images with blunt CP angle.
scoring system.^8 We have performed two sets of experiments. We calculated accuracy of the placement of the patches i.e. the images where costophrenic angle point was accurate enough to extract patch at the correct location. Based on visual assessment by a trained reader, the patch localization accuracy is 91%, 88%, and 92% for normal images, abnormal images with only blunt CP angle and abnormal images with blunt CP angle including other abnormalities, respectively.
CP angle measurement was done for left and right patches and then maximum among the two angles is used to assign CP angle to the image. In the box plot (Fig. 3), we see higher CP angle values for images for abnormal images with blunt costophrenic angle. Abnormal images tend to have large amount of fluid in pleural recess increasing the bluntness of CP angle. Images with only blunt CP angle usually have less fluid in the costophrenic recess making the CP angle measurement smaller. The average CP angles for normal, abnormal with only blunt CP angle and abnormal with blunt CP angle including other abnormalities are 71.6◦^ ± 22.9, 87.5◦^ ± 25.7, and 87.7◦^ ± 25.3 respectively. Clearly there is an overlap present in CP angle values between normal and abnormal images. Visual inspection shows that this is sometimes due to wrong placement of the patch or an incorrect lung segmentation mask. Incorrect location of the CP angle, hemidiaphragm and lateral chest wall points causes wrong CP angle calculation. To rectify this problem, a classifier could be trained to reclassify the patch for better mask generation. Also fitting a line to the contour points on hemidiaphragm and lateral chest wall points^6 could potentially improve the CP angle calculation. Example images of wrong CP angle calculation are shown in 5.
4. CONCLUSION
A technique for localization of the costophrenic recesses and costophrenic angle measurements has been presented and evaluated on a large data set of normal as well as abnormal CXRs. The result is a promising first step towards automated assessment of the presence of pleural fluid in the costophrenic recesses on chest radiographs.
Normal case, θright = 37.4◦, θlef t = 37.4◦
Abnormal case with only blunt CP angle, θright = 46.5◦, θlef t = 106.2◦
Abnormal case with blunt CP angle and other abnormalities, θright = 46.5◦, θlef t = 136.1◦
(a) (b) (c) (d) (e)
Figure 4. Costophrenic angle point is marked by a cross. Column (a) Right costophrenic recess patch segmentation, (b) Right costophrenic recess patch, (c) Original CXR image with outlined lung segmentation contour, (d) Left costophrenic recess patch, (e) Left costophrenic recess patch segmentation
Figure 5. Example images with wrong costophrenic angle measurement due to incorrect lung segmentation.
