Computer assisted detection and modelling of paediatric airway pathology from medical images
Benjamin Irving
5 September 2012
Navigate slides using left and right buttons
Clinical motivation
- TB still has a considerable effect of quality of life in developing countries
- Incidence of TB in South Africa 981 per 100 000 / year
- Especially difficult to detect in children
- Need for additional methods of detection
- Lymphadenopathy is commonly present in paediatric tuberculosis
- ≈ 90% - 100% of patients under 5 years
- Affects the airway
- Additional tests based on analysis of airway deformation
Aim
- Develop a classification system based on 3D airway shape deformation for detection of TB in 3D CT with the potential to extend to 2D radiographs
Thesis outline
- Segmentation
- Structure
- Registration and correspondence
- Obstruction detection
- Classification of airway pathology
- Shape model projection and X-ray segmentation
Airway segmentation
Coronal CT slice showing segmented airway regions
Morphological filtering a) region of an axial slice b) segmentation after coronal and sagittal filtering c) segmentation after axial filtering
Segmentation (Interactive Figure)
Mesh representation of airway segmentation
Click and drag figure to rotate. Scroll to zoom.
Click on figure and press 'm' to toggle mesh views.
Structure and skeletonisation
Airway segmentation and labelled skeleton. Colours are used to distinguish branches in the skeleton and a background CT slice is provided to demonstrate position
Correspondence
Generating surface landmarks in the region of interest on the airway surface
Registration (Interactive Figure)
Registration of a template mesh to each airway region using thin-plate-spline and closest point alignment
Slider for each transform
Move each slider to the right to perform template warp
Click and drag figure to rotate. Scroll to zoom.
Obstruction detection
Detecting and segmenting beyond points of obstruction
Classification of airway pathology (Interactive)
A single PCA mode of variation for the airway dataset
PCA mode 1
(+ λ → 0 → - λ)
Click and drag figure to rotate. Scroll to zoom.
Move the slider to show the variation along one statistical mode
Classification of airway pathology
A number of modes of variation are used to train a classifier to detect airway pathology
Classification of airway pathology
ROC curves for classification of paediatric TB from airway shape deformation
Projection of 3D model onto 2D surface
Projection of silhouette edge vertices onto a 2D surface.
Used to assist in the segmentatio of the airways in 2D radiographs
Challenges
- Development of the entire project from scratch
- Collection of data
- Collaborations and ethics
- Cases
- Novel dataset
- Segmentation gold standard
Contributions and achievments
- Extension of morphology based airway segmentation algorithms and application to paediatric airway segmentation
- A method to develop dense mesh correspondence between airway trees
- Detection of airway tree obstructions
- Application of PCA to variation of regions of the airway
- A CAD system to detect TB in CT from normal and abnormal airway shape variation
- Projection of the airway shape model to segment the airways in radiographs
(a big step towards 2D airway analysis)