Computer assisted detection of abnormal airway variation in CT scans related to paediatric tuberculosis
Graphical abstract
Introduction
Tuberculosis is still prevalent – particularly in developing countries. The WHO estimates the annual number of new cases of TB to be 343/100,000 in Africa, higher than any other region (WHO, 2012). Childhood tuberculosis represents a large proportion of TB cases and is particularly difficult to diagnose; the disease is confirmed in less than 40% of cases (Schaaf et al., 1995). While most TB in adults is confirmed from sputum samples, childhood TB has a low sputum bacilli count, and, therefore, the diagnosis is based on a number of factors, with imaging playing a major role. Other indicators include: contact with an infected adult (often difficult to determine); symptoms and signs (vague and common); and tuberculin skin test (positive test only indicates exposure and false negatives can be attributed to HIV or other immune disorders). Each of these tests is imperfect but a combination is used to diagnose TB (Gie, 2003). Recent studies have highlighted increasing clinical interest and the need for further research into the diagnosis and treatment of paediatric TB (Sandgren et al., 2012).
Lymph node involvement is a key indicator of childhood TB, and as the lymph nodes enlarge the airways are compressed or deformed. This is because paediatric patients have smaller airways with less well developed cartilage – predisposing them to compression. This is known as lymphobronchial TB and is a current clinical focus in childhood TB research (Lucas et al., 2012, Goussard et al., 2013). Airway involvement due to lymphadenopathy is relatively common in children (between 29% (Andronikou et al., 2004) and 38% (Theart et al., 2005) of cases). This deformation can occur due to one enlarged lymph node, or compression can be due to an enlarged lymph node on both sides or a lymph node and a vessel, which can be used as an indicator of TB (Goussard and Gie, 2007, Andronikou et al., 2004). Fig. 1 shows a MinIP projection with airway stenosis caused by lymphadenopathy in a paediatric pulmonary TB patient. The most affected airways include the bronchus intermedius (BI), left main bronchus (LMB), trachea and right main bronchus (RUL) (Lucas et al., 2012, Goussard et al., 2013).
Determining airway involvement and cause is important for the treatment of paediatric TB and other diseases affecting the airways (Andronikou et al., 2013). Bronchoscopy is the “gold standard” for determining airway involvement but is invasive, general anaesthesia is often required and the external cause of the airway involvement cannot be seen. Recent studies suggest that CT with volume rendering can be used as an alternative to bronchoscopy (du Plessis et al., 2009) – offering the benefits of bronchoscopy while also allowing visualisation of the external cause of the airway involvement. This method allows manual measurements of airway cross-sections from the 3D rendering of the region, but requires considerable manual interaction in the form of setting thresholds, viewing parameters, and manual assessment of the airways, and a more automated approach to monitor airway involvement would be beneficial.
Thus, there is considerable value in automatically detecting airway involvement from lymphadenopathy to assist in the detection and assessment of paediatric patients with tuberculosis (and potentially other diseases). In this study we developed a method we call the local airway point distribution model (LA-PDM) to assess normal and pathological variation in local regions of the airway. Point distribution models are effective for capturing variation that is more complex than airway narrowing – where more complex variation is related to airway involvement from lymphadenopathy.
Section 2 discusses current methods used for airway analysis and Section 3 outlines our proposed LA-PDM method. This method performs a 3D airway segmentation using chest CT images, and then a branch labelling of the segmented airway (Section 3.1). Surface point correspondence is developed between the dataset of segmentations (Section 3.3 Mesh alignment step 1: thin-plate-spline warp, 3.4 Mesh alignment step 2: local alignment) and used to create point distribution models for local regions of the airway (Section 3.6). The LA-PDM can then be used to distinguish normal and abnormal variation in local airway regions of a new CT image. Section 4 applies the LA-PDM to a 90 patient test dataset of paediatric chest CT scans. The results of the LA-PDM on the test set is shown in Section 5 and compared to simpler branch diameter based features, and show promising results for detection of airway pathology related to paediatric TB.
Section snippets
Airway shape analysis
A number of studies have developed methods that automate branch diameter measurements as well as the broncho-aterial ratio from segmented adult airways (Kiraly et al., 2008, Tschirren et al., 2005, Palágyi et al., 2006). Applications include detection of chronic obstructive pulmonary disease (COPD) and asthma identification (Petersen et al., 2010, Fetita et al., 2010, Wiemker et al., 2004). These methods are useful but rely on the identification of pathology from variation in local airway
Local airway point distribution models for abnormality detection
Fig. 2 shows our proposed method that uses a PDM of local regions of the airway to detect airway abnormalities. Airways are segmented and the branches are labelled. Surface mesh correspondence is developed using a reference airway, and PDMs with trained classifiers are used to detect airway abnormalities.
Application of the LA-PDM to clinical cases
This section describes the application of the LA-PDM method to the paediatric chest CT dataset.
Detection of cases with airway abnormalities
The CAD pipeline that we have developed is able to detect airway deformation related to paediatric TB. The following steps are used for detection of an unseen case: a chest CT of a paediatric patient with suspected TB is acquired, the airway is automatically segmented, the airway centreline and branch points are identified, and landmarks are placed on the airway surface using the methods described in Section 3.1 Segmentation and branch labelling, 3.2 Surface point projection. The airway
Discussion
The LA-PDM method introduced in this paper can accurately distinguish between airway involvement (from paediatric pulmonary TB) and normal airways by examining regions of the airway likely to be affected by lymphadenopathy (AUC of and ). The LA-PDM derived features show more promise than features derived from the airway diameter, which is probably due to the ability to represent more complex variation in feature space. However, we are not aware of any previous
Acknowledgements
The authors would like to thank the Commonwealth Scholarship Commission (CSC) for funding this research.
References (42)
Shape and the information in medical images: a decade of the morphometric synthesis
Comput. Vis. Image Underst.
(1997)- et al.
Active shape models-their training and application
Comput. Vis. Image Underst.
(1995) - et al.
A comparison of the accuracy of statistical models of prostate motion trained using data from biomechanical simulations
Prog. Biophys. Mol. Biol.
(2010) - et al.
Quantitative analysis of pulmonary airway tree structures
Comput. Biol. Med.
(2006) - et al.
Assessment and stenting of tracheal stenosis using deformable shape models
Med. Image Anal.
(2011) - et al.
Statistics with Confidence
(2005) - et al.
Modern imaging of tuberculosis in children: thoracic, central nervous system and abdominal tuberculosis
Pediatr. Radiol.
(2004) - et al.
CT scanning for the detection of tuberculous mediastinal and hilar lymphadenopathy in children
Pediatr. Radiol.
(2004) - et al.
Technical developments in postprocessing of paediatric airway imaging
Pediatr. Radiol.
(2013) Principal warps: thin-plate splines and the decomposition of deformations
IEEE Trans. Pattern Anal. Mach. Intell.
(1989)