Embed Size (px)
Citation preview
MODULE II
Semi-Automatic Segmentation
RSNA 2016 ITK-SNAP Course
Semi-Automatic Segmentation
Segmentation Algorithm
User: quickly identifies structure of interest and sets parameters
Semi-Automatic Segmentation
Segmentation Algorithm
User: quickly identifies structure of interest and sets parameters
Semi-Automatic Segmentation
Segmentation Algorithm
User: quickly identifies structure of interest and sets parameters
Semi-Automatic Segmentation
Segmentation Algorithm
User: quickly identifies structure of interest and sets parameters
How does it work?
• Pre-segmentation Phase
• Identify parts of image as foreground and background
• Active Contour Phase
• Place seeds inside the structure of interest
• Grow the seeds to fill the structure of interest
Pre-segmentation Phase
Images Region of Interest
Pre-segmentation
Speed Image
Foreg
round
Backg
round
Uncert
ain
Pre-segmentation Modes
Thresholding
Classification
Pre-segmentation Modes
Thresholding
Classificationsets t
hresholds
Pre-segmentation Modes
Thresholding
Classificationsets t
hresholds
Pre-segmentation Modes
Thresholding
Classificationsets t
hresholds
paints examples
Pre-segmentation Modes
Thresholding
Classificationsets t
hresholds
paints examples
Active Contour Evolution driven by Forces
Image Force: outward over foreground, inward over background
Smoothing Force: strongest in places of high curvature
Active Contour Evolution
Active Contour Evolution
Active Contour Evolution
Active Contour Evolution
places seeds
Live demo of semi-automatic segmentation
Newest ITK-SNAP Functionality (late 2016)
Manual and automatic registrationMorphological interpolation
2
1 Introduction
Proliferation of automatic segmentation algorithms is on the rise, but their performance is measured againsta manually created ‘gold standard’ segmentation. Additionally, manual segmentation is sometimes requiredfor other purposes.
As anyone who has ever done manual image segmentation can tell you, it becomes monotonous very quickly.This is especially true for high-resolution 3D images, i.e. for 3D images with thin ‘slices’ (because thereare many of them). The classic ‘cure’ for this is inter-slice interpolation (Fig. 1). Namely, in regions wherethe anatomy is slowly changing and adjacent slices are quite similar, some slices can be skipped by a humansegmenter and then interpolated by an algorithm.
Figure 1: Examples of inputs (left) and outputs (right). First row: two labels, manual segmentation on coronal slices
only. Second row: four labels, manual segmentation along all three principal axes of the body.
Latest version available at the Insight Journal [ http://hdl.handle.net/10380/1338]Distributed under Creative Commons Attribution License
2
1 Introduction
Proliferation of automatic segmentation algorithms is on the rise, but their performance is measured againsta manually created ‘gold standard’ segmentation. Additionally, manual segmentation is sometimes requiredfor other purposes.
As anyone who has ever done manual image segmentation can tell you, it becomes monotonous very quickly.This is especially true for high-resolution 3D images, i.e. for 3D images with thin ‘slices’ (because thereare many of them). The classic ‘cure’ for this is inter-slice interpolation (Fig. 1). Namely, in regions wherethe anatomy is slowly changing and adjacent slices are quite similar, some slices can be skipped by a humansegmenter and then interpolated by an algorithm.
Figure 1: Examples of inputs (left) and outputs (right). First row: two labels, manual segmentation on coronal slices
only. Second row: four labels, manual segmentation along all three principal axes of the body.
Latest version available at the Insight Journal [ http://hdl.handle.net/10380/1338]Distributed under Creative Commons Attribution License
Newest ITK-SNAP Functionality (late 2016)
Manual and automatic registrationMorphological interpolation
2
1 Introduction
Proliferation of automatic segmentation algorithms is on the rise, but their performance is measured againsta manually created ‘gold standard’ segmentation. Additionally, manual segmentation is sometimes requiredfor other purposes.
As anyone who has ever done manual image segmentation can tell you, it becomes monotonous very quickly.This is especially true for high-resolution 3D images, i.e. for 3D images with thin ‘slices’ (because thereare many of them). The classic ‘cure’ for this is inter-slice interpolation (Fig. 1). Namely, in regions wherethe anatomy is slowly changing and adjacent slices are quite similar, some slices can be skipped by a humansegmenter and then interpolated by an algorithm.
Figure 1: Examples of inputs (left) and outputs (right). First row: two labels, manual segmentation on coronal slices
only. Second row: four labels, manual segmentation along all three principal axes of the body.
Latest version available at the Insight Journal [ http://hdl.handle.net/10380/1338]Distributed under Creative Commons Attribution License
2
1 Introduction
Proliferation of automatic segmentation algorithms is on the rise, but their performance is measured againsta manually created ‘gold standard’ segmentation. Additionally, manual segmentation is sometimes requiredfor other purposes.
As anyone who has ever done manual image segmentation can tell you, it becomes monotonous very quickly.This is especially true for high-resolution 3D images, i.e. for 3D images with thin ‘slices’ (because thereare many of them). The classic ‘cure’ for this is inter-slice interpolation (Fig. 1). Namely, in regions wherethe anatomy is slowly changing and adjacent slices are quite similar, some slices can be skipped by a humansegmenter and then interpolated by an algorithm.
Figure 1: Examples of inputs (left) and outputs (right). First row: two labels, manual segmentation on coronal slices
only. Second row: four labels, manual segmentation along all three principal axes of the body.
Latest version available at the Insight Journal [ http://hdl.handle.net/10380/1338]Distributed under Creative Commons Attribution License
Please work on exercises 3 and 4
