-
image segmentation
- partitioning (or labelling) of foreground from background
- typically an image of 0s and 1s
- representation can be of either boundary or of region
-
methods of segmentation (5)
- global thresholding: any pixel greater than threshold gets specific label
- region growing: graph theory approach, breadth first search
- connectivity
- unsupervised ML: k-means clustering
- supervised ML: training by painting metaphor
-
Image filtering
- smoothing out an image while preserving edges
- noise vs. real features (edges) is the key struggle
-
image filtering: taking derivatives makes it ____
worse (finite differences)
-
image filtering: you can average it out
convolution, canny edge detection
-
image filtering: you can statistically remove outliers
alpha-trimmed mean
-
image filtering: selectively blur out
anisotropic diffusion
-
image filtering: why?
- filter out unwanted noise
- detect desireable features
-
image filtering techniques
- convolution and smoothing: continuous, discrete, Gaussian smoothing
- edge detection: canny edge detector, Laplacian of Gaussian Edge Detector
- non-linear noise reduction: Alpha-trimmed mean, perona-malik anisotropic diffusion
-
image registration
aligning images so that the correspondences between them can be seen more easily
-
image registration: coordinate systems - adding a dimension
- add a dimension so every transformation is a matrix multiplication
- translation is usually not a multiplication
-
image registration: rigid transformation
- preserves angles, distances, and handedness
- hard anatomy (head)
- bones
-
image registration: affine transformation
- preserves co-linearity, parallel lines
- scaling due to size changes
- shear due to body leaning
-
image registration: perspective transformation
- preserved incidence, cross ratios
- 2D/3D representation
- pictures from two different perspectives
-
image registration: curved transformation
- surgical changes
- soft tissue deformation
- changes in body position
-
Example of image analysis pipeline: diagnose subtype of brain cancer
- divide huge pathology image into sub-images
- feature reduction and clustering to identify similar regions
- PCA to reduct feature space, k-means clustering on PCs with largest eigenvalues
- tile selection and deep feature extraction
- elastic net modeling and weighted voting to predict subtype of brain cancer
-
example of image analysis pipeline: diagnosis to predict disease progression in macular degeneration
- reasoning with quantitative features in longitudinal images
- use optical coherence images of retina - non-invasive 3D imaging technique
- automated segmentation: recognize and circumscribe debris in eyes on images
- quantitative image feature extraction: objective features that characterize debris regions
- statistical modeling/machine learning: predict AMD progression from image features
-
example of image analysis pipeline: reasoning with semantic image features - thyroid nodule diagnosis
- goal: develop decision support application to predict diagnosis of benign vs. malignant using ultrasound images of thyroid gland
- Bayesian network: given set of findings what is the probability that disease is present?
- probability of ultrasound features come from expert knowledge and literature
-
example of image analysis pipeline: image-based reasoning with anatomical knowledge - predict consequences of penetrating injury
- images lack explicit information about their contents (anatomical structures, tissue properties, physiological status of organs)
- image segmentation to identify anatomic structures
- link image regions to ontology of anatomy/physiology
- ontology provides artierial supply to organs, typology of diseases
- link ontology knowledge using semantic annotation (via AIM)
- computer reasoning with image regions and anatomy
-
Image segmentation: distance transform
- useful to know how far each pixel is from the object boundary
- applications: navigation through organs without bumping into walls
- analysis of shape similarity
- determination of geometrically "special" points
- input: binary image
- output: grayscale distance map image
|
|
