lec8: medical image segmentation (ii) (region growing/merging)

MEDICAL IMAGE COMPUTING (CAP 5937)

LECTURE 8: Medical Image Segmentation (II)(Region Growing/Merging)

Dr. Ulas BagciHEC 221, Center for Research in Computer Vision (CRCV), University of Central Florida (UCF), Orlando, FL [email protected] or [email protected]

1SPRING 2017

Outline• Region Growing algorithm• Homogeneity Criteria• Split/Merge algorithm• Examples from CT, MRI, PET• Limitations

2

Region Based Segmentation AlgorithmsRegion:A group of connected pixelswith similar properties

Closed boundaries

Computation of regions is based on similarity

Regions may correspond toObjects in a scene or partsof objects

Spatial proximity + similarity

3

Region Growing/Merging• Perhaps, one of the simplest approaches among region

based methods.

4

1. Select a seed point/points2. Define a growth criteria3. Joint all voxels connected to the seed

that follow the growth criteria4. Stop when no adjacent voxel agree with

the growth criteria

4-connectivity 8-connectivity

Example: Fat Segmentation in MRI

5

T1 weighted MRIRed: visceral fatPurple: liverBlue: Subcutaneous fat

Credit: Dougherty, G., Medica Image Processing.

Region-based Segmentation6

3D segmented two nodules are shown.Credit: Dehmeshki, et al. IEEE TMI 2008.

Red: interactive RG, white: manual segm. of lung regions pertaining to rabbits with infectionsCredit: Bagci et al. EJNMMI Research 2013

Simplest RG Pseudo-Code

7

Pictorial Illustration

8

1. Choose the seed pixel


9

1. Choose the seed pixel2. Check the neighboring pixels and add them to the region if

they are similar to the seed


10

1. Choose the seed pixel2. Check the neighboring pixels and add them to the region if

they are similar to the seed3. Repeat step 2 for each of the newly added pixels; stop if no

more pixels can be added

|neighboring pixels� seed| < Threshold

Algorithm Illustration

11

Split and Merge Algorithm• Unlike RG, region splitting starts with the whole image as a single region

and subdivides it into subsidiary regions recursively while a condition of homogeneity is not satisfied.

12



• Region merging is the opposite of splitting, and works as a way of avoiding over-segmentation

13



• Region merging is the opposite of splitting, and works as a way of avoiding over-segmentation

14

Drawback of Split ?

15

Drawback of Split ?• the final image partition may contain adjacent regions with

identical properties.• Solution?

16

Drawback of Split ?• the final image partition may contain adjacent regions with

identical properties.• Solution?

– REGION MERGING

17

Split and Merge Algorithm• When a homogeneous region is created, its neighboring

regions are checked and the newly created region is merged with an existing one if they have identical properties.

• If the similarity criteria are met by more than one adjacent region, the new region is merged with the most similar one.

18

Muscle/Bone Separation in CT

19


20


21


22


23

Virtual Endoscopy/Colonoscopy - CT24

(a) Original CT imagethrough the colon. (b) A 3D RG initiated from a seed point. (c) 3D rendering of thesegmented colon.

Credit: P. Seutens

Adaptive Parameterization of RG• Literature is vast!

25

Adaptive Parameterization of RG• Literature is vast!• Ex: Tumor segmentation from PET images

26

Adaptive Parameterization of RG• Literature is vast!• Ex: Tumor segmentation from PET images

T: threshold, mean of R0 à mean intensity of current grown region

27

Adaptive Parameterization of RG

28

Credit: Hua Li et alMedPhys 2008

a. Original imageb. Threshold/Volume curvec. ROId. Thresholdinge. Adaptive RG

Left Ventricular Function in Cardiac CT• The gold standard for the functional evaluation of the left

ventricle (LV) is magnetic resonance imaging (MRI)

29


ventricle (LV) is magnetic resonance imaging (MRI)• Contrast-enhanced retrospectively electrocardiogram (ECG)-

gated multi-slice spiral computed tomography (MSCT) has been accepted as an efficient noninvasive tool for the detection of coronary artery stenosis, providing good sensitivity and specificity

30


ventricle (LV) is magnetic resonance imaging (MRI)• Contrast-enhanced retrospectively electrocardiogram (ECG)-

gated multi-slice spiral computed tomography (MSCT) has been accepted as an efficient noninvasive tool for the detection of coronary artery stenosis, providing good sensitivity and specificity

• Assessment of the LV volume at end diastole and end systole is used for quantification of function!

31

Left Ventricular Function in Cardiac CT

32

Endocardial contours in systole Endocardial contours in diastole

Credit: muhlenbruch, et al. Eur Rad 2006

Homogenous region and high contrasthelps region growing to be suitable for this case!

CT Lung/Soft Tissue Separation• Seed > -200 HU

33

Airway Segmentation (Basic RG)

34

Slicer Implementation of RG and Airway Segmentation

35

Airway Segmentation (Basic RG)

36

Credit: Nardelli et alBioMedOnline 2015

Bone Vanishing Disease - CT

37


38

(credit: Papadakis, G.Z., NIH


39

(credit: Papadakis, G.Z., NIH

Controlling the homogeneityparameter (similarity criteria) in RGis extremely difficult due to largevariation in bone density shifting towardslower densities

-78 HU

307 HU

Limitations of Region Growing1) Objects in medical images are often non-homogeneous

40


2) Poor contrast causes leakages

41


2) Poor contrast causes leakages

3) Noise

42

Summary• Region Growing based segmentation algorithms

– Still in use for many applications in radiology• Lung, bone, homogenous (isolated ) tumors, …

– Either terminal use or in-between step– Similarity criteria: homogenous regions

43

Slide Credits and References• Jayaram K. Udupa, MIPG of University of Pennsylvania, PA.• P. Suetens, Fundamentals of Medical Imaging, Cambridge

Univ. Press.

• Adams and Bischof, IEEE PAMI 1994.• Zhu and Yuille, IEEE PAMI 1996.• Hu et al, IEEE TMI 2001.• Dehmeshki et al IEEE TMI 2008.• Nardelli et al BioMedOnline 2015• Dougherty, G., Medical Image Processing• Hua Li et al MedPhys 2008

44

lec8: medical image segmentation (ii) (region growing/merging)

Science