SPRING 2016
MEDICAL IMAGE COMPUTING (CAP 5937)
LECTURE 7: Medical Image Segmentation (I) (Radiology Applications of Segmentation, and Thresholding)
Dr. Ulas Bagci HEC 221, Center for Research in Computer Vision (CRCV), University of Central Florida (UCF), Orlando, FL 32814.
[email protected] or
[email protected]
1
2
Outline • Introduction to Medical Image Segmentation, type of segmentation methods, and definitions – Recognition & Delineation
• Simplest Segmentation Method(s): Thresholding – – –
Otsu Thresholding Parametric Method PET Image Thresholding Methods •
ITM (Iterative Thresholding Method)
3
Motivation for Image Segmentation In the last 20 years the computer vision and medical imaging communities have produced a number of useful algorithms for localizing object boundaries in images.
4
Motivation for Image Segmentation • • • • • • • • • • •
Content based image retrieval Machine Vision Medical Imaging applications (tumor delineation,..) Object detection (face detection,…) 3D Reconstruction Object/Motion Tracking Object-based measurements such as size and shape Object recognition (face recognition,…) Fingerprint recognition, Video surveillance …
5
Segmentation Tools in Radiology Applications • 3D views to visualize structural information and spatial anatomic relationships is a difficult task, which is usually carried out in the clinician’s mind.
6
Segmentation Tools in Radiology Applications • 3D views to visualize structural information and spatial anatomic relationships is a difficult task, which is usually carried out in the clinician’s mind. • Image-processing tools provide the surgeon with interactively displayed 3D visual information.
7
Segmentation Tools in Radiology Applications
Credit: Kaus, et al. Radiology 2001.
8
Segmentation Tools in Radiology Applications • Determination of the volumes of abdominal solid organs and focal lesions has great potential importance (liver, spleen, …). • Monitoring the response to therapy and the progression of neoplastic disease and preoperative examination of living liver donors are the most common clinical applications of volume determination.
(credit: Farraher, et al. Radiology 2005)
9
Segmentation Tools in Radiology Applications • Gross Tumor Volume in CT/MRI • Metabolic Tumor Volume in PET/SPECT/ – Surgery/Therapy Planning • Planning Tumor Volume (PTV)
– Tumor characterization
• Texture Extraction requires segmentation to be done • Shape analysis
Credit: Manniesing, et al, Radiology 2008
Segmentation Tools in Radiology Applications • There is a strong interest in automatic and reproducible techniques for detection and quantification of vascular disease • A first step toward an effective vessel analysis tool is segmentation of the vasculature. axial
MIP: maximum intensity Projection image of cerebral vessels (in CTA)
coronal
sagittal
10
11
Segmentation Tools in Radiology Applications • MR volumetry of the hippocampus can help distinguish patients with AD (Alzheimer’s Disease) from elderly controls with a high degree of accuracy (80%–90%).
12
Segmentation Tools in Radiology Applications • MR volumetry of the hippocampus can help distinguish patients with AD (Alzheimer’s Disease) from elderly controls with a high degree of accuracy (80%–90%).
amygdala Credit: Colliot et al, Radiology 2008.
hippocampus
13
Image Segmentation Definition: Partitioning a picture/image into distinctive subsets is called segmentation.
14
Image Segmentation Definition: Partitioning a picture/image into distinctive subsets is called segmentation.
Segmentation of an image entails the division or separation of the image into regions of similar attribute.
15
Image Segmentation Definition: Partitioning a picture/image into distinctive subsets is called segmentation.
Segmentation of an image entails the division or separation of the image into regions of similar attribute.
The most basic attributes: -intensity -edges -texture -other features…
16
Image Segmentation Definition: Partitioning a picture/image into distinctive subsets is called segmentation.
Purpose:
To extract object information and represent this as a hard/fuzzy geometric structure.
Recognition:
Determining the object’s whereabouts in the scene. (humans > computer)
Delineation:
Determining the object’s spatial extent and composition in the scene. (computers > humans)
17
Recognition - Example
Model is induced
No Model is induced (slice credit: J. Kim et al, Signal Processing 2007)
18
Approaches to Recognition • Model-based • Knowledge-based • Atlas-based
-
Non-interactive
• Human-assisted
-
Interactive
19
Approaches to Recognition • Model-based • Knowledge-based • Atlas-based
-
Non-interactive
• Human-assisted
-
Interactive
- They all originate from human knowledge. - Their relative efficacy is unknown.
20
Approaches to Delineations pI (purely image-based) approaches • Rely mostly on information available in the given image only. • Recognition: manual
21
Approaches to Delineations pI (purely image-based) approaches • Rely mostly on information available in the given image only. • Recognition: manual SM (shape model-based) approaches • Employ models to codify object family shape info. • Recognition: model-based/manual
22
Approaches to Delineations pI (purely image-based) approaches • Rely mostly on information available in the given image only. • Recognition: manual SM (shape model-based) approaches • Employ models to codify object family shape info. • Recognition: model-based/manual Hybrid approaches • Combine among pI and SM approaches. • Recognition: model-based, automatic.
23
Classification of Methods Boundary-based (BpI): • optimum boundary • active boundary • live wire • level sets
24
Classification of Methods Boundary-based (BpI): • optimum boundary • active boundary • live wire • level sets Region-based (RpI): • clustering – kNN, CM, FCM • graph cut • fuzzy connectedness • MRF • watershed • optimum partitioning • (Mumford-Shah)
25
Classification of Methods Boundary-based (BpI): • optimum boundary • active boundary • live wire • level sets Region-based (RpI): • clustering – kNN, CM, FCM • graph cut • fuzzy connectedness • MRF • watershed • optimum partitioning • (Mumford-Shah)
SM Approaches • manual tracing • live wire • active shape/appearance • M-reps • atlas-based
26
Classification of Methods Boundary-based (BpI): • optimum boundary • active boundary • live wire • level sets Region-based (RpI): • clustering – kNN, CM, FCM • graph cut • fuzzy connectedness • MRF • watershed • optimum partitioning • (Mumford-Shah)
SM Approaches • manual tracing • live wire • active shape/appearance • M-reps • atlas-based
Hybrid Approaches • • • • • •
BpI + BpI RpI + RpI BpI + RpI BpI + SM RpI + SM SM + SM
27
Classification of Methods pI Approaches + Where image info is good, accuracy is good; - Bad where it is poor/absent; - Need recognition help; + Can determine degree of match of model to image well; - Lack obj shape & geographic info;
28
Classification of Methods SM Approaches - Even where image info is good, accuracy suffers; + Where bad, model helps; + Can help in recognition; - Need best match info;
+ Good models embody obj shape & geographic info;
29
Purely Image Based Segmentation Methods
30
Thresholding – Simple Segmentation • Image binarization – mapping a scalar image I into a binary image J
J(x, y) =
(
0 1
if I(x, y) < T otherwise.
31
Thresholding – Simple Segmentation • Image binarization – mapping a scalar image I into a binary image J
J(x, y) =
(
0 1
if I(x, y) < T otherwise.
32
Thresholding – Simple Segmentation Brighter objects
Darker objects
33
Thresholding – Simple Segmentation Brighter objects
Darker objects
DIFFICULTIES 1. The valley may be so broad that it is difficult to locate a significant minimum 2. Number of minima due to type of details in the image 3. Noise 4. No visible valley 5. Histogram may be multi-modal
34
Example: CT Scan
35
Example: CT Scan
36
Example: CT Scan
37
Example: CT Scan
38
Example: CT Scan
39
Thresholding Methods • • • • • • • • • • •
Huang Intermode Isodata Li MaxEntropy Mean MinError Otsu Percentile RenyiEntropy Moments
40
Thresholding Methods • • • • • • • • • • •
Huang Intermode Isodata Li MaxEntropy Mean MinError Otsu Percentile RenyiEntropy Moments
41
Thresholding Methods • • • • • • • • • • •
Huang Intermode Isodata Li MaxEntropy Mean MinError Otsu (non-parametric) Percentile RenyiEntropy Moments
PET Imaging Fixed Thresholding Adaptive Thresholding Iterative Thresholding
42
Otsu Thresholding • Definition: The method uses the grey-value histogram of the given image I as input and aims at providing the best threshold in the sense that the “overlap” between two classes, set of object and background pixels, is minimized (i.e., by finding the best balance).
43
Otsu Thresholding • Definition: The method uses the grey-value histogram of the given image I as input and aims at providing the best threshold in the sense that the “overlap” between two classes, set of object and background pixels, is minimized (i.e., by finding the best balance). • Otsu’s algorithm selects a threshold that maximizes the 2 between-class variance b . In the case of two classes, 2 b
= P1 (µ1
µ)2 + P2 (µ2
µ)2 = P1 P2 (µ1
µ2 ) 2
44
Otsu Thresholding • Definition: The method uses the grey-value histogram of the given image I as input and aims at providing the best threshold in the sense that the “overlap” between two classes, set of object and background pixels, is minimized (i.e., by finding the best balance). • Otsu’s algorithm selects a threshold that maximizes the 2 between-class variance b . In the case of two classes, 2 b
= P1 (µ1
µ)2 + P2 (µ2
µ)2 = P1 P2 (µ1
µ2 ) 2
• where P1 and P2 denote class probabilities, and μi the means of object and background classes.
45
Otsu Thresholding • Definition: The method uses the grey-value histogram of the given image I as input and aims at providing the best threshold in the sense that the “overlap” between two classes, set of object and background pixels, is minimized (i.e., by finding the best balance).
P1 =
u X
u
p(i)
ı=0
P2 =
GX max
ı=u+1
p(i)
u
46
Otsu Thresholding • Definition: The method uses the grey-value histogram of the given image I as input and aims at providing the best threshold in the sense that the “overlap” between two classes, set of object and background pixels, is minimized (i.e., by finding the best balance).
P1 =
u X
p(i)
µ1 =
P2 =
ı=u+1
p(i)
ip(i)/P1
ı=0
ı=0
GX max
u X
µ2 =
GX max
ip(i)/P2
ı=u+1 CLASS MEANS
47
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
48
Otsu Thresholding-Algorithm
2 b
P1 cI (u)
= P1 (µ1
µ)2 + P2 (µ2
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
µ)2 = P1 P2 (µ1
µ2 ) 2
49
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
50
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
51
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
52
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
53
Otsu Thresholding-Algorithm
P1 cI (u)
P2 1
cI (u)
c indicates cumulative histogram, and P1 and P2 can be approximated well with cumulative density function.
optimal
54
Parametric Method for Optimal Thresholding • Assuming again a two-class problem and assuming that the distribution of gray levels for each class can be modeled by a normal distribution with mean and variance
55
Parametric Method for Optimal Thresholding • Assuming again a two-class problem and assuming that the distribution of gray levels for each class can be modeled by a normal distribution with mean and variance • the overall normalized intensity histogram can be written as the following mixture probability density function:
56
Parametric Method for Optimal Thresholding • Assuming again a two-class problem and assuming that the distribution of gray levels for each class can be modeled by a normal distribution with mean and variance • the overall normalized intensity histogram can be written as the following mixture probability density function:
where P1 and P2 are class probabilities. The optimal threshold (T) can be found as solving the quadratic equation à
57
Parametric Method for Optimal Thresholding
58
Parametric Method for Optimal Thresholding
In case, variances of both classes are equal, then->
59
Parametric Method for Optimal Thresholding
In case, variances of both classes are equal, then->
60
Thresholding methods for PET Image Segmentation • Due to the nature of PET images (i.e., low resolution with high contrast), thresholding-based methods are suitable – because the local or global intensity histogram usually provides a sufficient level of information for separating the foreground (object of interest) from the background. (Foster, Bagci, et al., CBM 2014)
61
Thresholding methods for PET Image Segmentation • Due to the nature of PET images (i.e., low resolution with high contrast), thresholding-based methods are suitable – because the local or global intensity histogram usually provides a sufficient level of information for separating the foreground (object of interest) from the background. (Foster, Bagci, et al., CBM 2014)
Fixed Thresholding
Adaptive Thresholding
Iterative Thresholding
62
Fixed Thresholding Methods
• Due to the nature of PET images (i.e., low resolution with high contrast), thresholding-based methods are suitable – because the local or global intensity histogram usually provides a sufficient level of information for separating the foreground (object of interest) from the background. (Foster, Bagci, et al., CBM 2014)
63
Thresholding methods for PET Image Segmentation • Due to the nature of PET images (i.e., low resolution with high contrast), thresholding-based methods are suitable – because the local or global intensity histogram usually provides a sufficient level of information for separating the foreground (object of interest) from the background. (Foster, Bagci, et al., CBM 2014)
Fixed Thresholding
Adaptive Thresholding
Phantom Based
Iterative Thresholding
Image Quality metrics based
Adaptive Thresholding
64
65
Thresholding methods for PET Image Segmentation • Due to the nature of PET images (i.e., low resolution with high contrast), thresholding-based methods are suitable – because the local or global intensity histogram usually provides a sufficient level of information for separating the foreground (object of interest) from the background. (Foster, Bagci, et al., CBM 2014)
Fixed Thresholding
Adaptive Thresholding
Phantom Based
Iterative Thresholding
Image Quality metrics based
66
Iterative Thresholding Method (ITM) S/B: Source to background ratio. The method is based on calibrated threshold-volume curves at varying S/B ratio acquired by phantom measurements using spheres of known volumes.
67
Iterative Thresholding Method (ITM) S/B: Source to background ratio. The method is based on calibrated threshold-volume curves at varying S/B ratio acquired by phantom measurements using spheres of known volumes.
68
Iterative Thresholding Method (ITM) S/B: Source to background ratio. The method is based on calibrated threshold-volume curves at varying S/B ratio acquired by phantom measurements using spheres of known volumes. The measured S/B ratios of the lesions are then estimated from PET images, and their volumes are iteratively calculated using the calibrated S/B-threshold-volume curves
69
Iterative Thresholding Method (ITM) S/B: Source to background ratio. The method is based on calibrated threshold-volume curves at varying S/B ratio acquired by phantom measurements using spheres of known volumes. The measured S/B ratios of the lesions are then estimated from PET images, and their volumes are iteratively calculated using the calibrated S/B-threshold-volume curves The resulting PET volumes are then compared with the known sphere volume and CT volumes of tumors that served as gold standards.
70
ITM Example Result on PET Images/Lung
71
Another Example for PET Thresholding
ITM for tumor segmentation/FDG PET
72
Another Example for PET Thresholding
73
Further Thresholding Example – CT Bones
74
Further Thresholding Example – CT Bones
75
Head-Neck CT – Thresholding for Skull Modeling
Segmentation of the skull and the mandibula in CT images using thresholding. (a) Original CT image of the head. (b) Result with a threshold value of 276 Hounsfield units. The segmented bony structures are represented in color. (c) 3D rendering of the skull shows a congenital growth deficiency of the mandibula in this 8-year-old patient. This information was used preoperatively to plan a repositioning of the mandibula. (Slice Credit: P.Seutens)
76
Multiple Thresholds – MRI Thresholding
Thresholding can be done interactively and separates the image into different regions. Valleys in the histogram indicate potentially useful threshold values Credit: Toeonies, K.
77
Summary of today’s lecture • Introduction into the Medical Image Segmentation • Recognition and Delineation concepts in Segmentation • Simplest Segmentation method: Thresholding – Otsu – Parametric method for optimal thresholding – PET Image thresholding • ITM, fixed thresholding, etc.
78
Slide Credits and References • Jayaram K. Udupa, MIPG of University of Pennsylvania, PA. • P. Suetens, Fundamentals of Medical Imaging, Cambridge Univ. Press. • • • • • • •
Foster, B., et al. CBM, Review paper, 2014. Kaus, et al. Radiology 2001. Toeonies, K., Medical Image Analysis. Farraher, et al., Radiology 2005 Zaidi, H., Quantitative Analysis in Nuclear Medicine Imaging. Bailey et al. Positron Emission Tomography, Springer. Dawood, M., et al. Correction Techniques in Emission Tomography