Adaptive Clustering in Medical Image Segmentation as a Diagnostic Tool

1
Adaptive Clustering in Medical Image Segmentation
as a Diagnostic Tool

• Sunanda Mitra and Shuyu Yang
• Computer Vision and Image Analysis Laboratory
• Department of Electrical and Computer Engineering
  Texas Tech University
• Sunanda.Mitra_at_coe.ttu.edu
• This work was partially supported by NSF AWARD
  9980296 CRCD Machine Learning A
  Multidisciplinary Computer Engineering Graduate
  Program and a seed funding from the National
  Library of Medicine.

2
Computer Vision and Image Analysis Laboratory
Sunanda Mitra Thomas F. Krile Tanja Karp Brian
Nutter Shuyu Yang
Faculty
M.S. students
Ye Linning Sandhya Borra Yaman Hossain Yeshwanth
Srinivasan Kayla Kepley Dana Hernes Bhakti
Tulpule Prateek Srivastava Hari Nayar Jie
Yin Zhenpeng Feng Chris Caceres VS Sampath Amit
Mane
Ph.D. Students
Jiang Ling Guo Kaan Aydin Philip King Andrew
Patterson Roopesh Kumar Sujit Joshi Ryan Casey Y.
Sriraja
3
Drawbacks of current imaging based diagnostic
tools

1. Subjective Variability in Interpretation
2. Lack of quantitative documentation
3. Lack of Standardization
4. Time-consuming manual segmentation of regions of
   interest
5. Absence of medical specialists in all locations

4
Medical image segmentation techniques

• Segmentation techniques developed for medical
  images are non-universal, and image modality and
  application specific.
• Compared to other segmentation approaches,
  clustering techniques can be made adaptive and
  are relatively simple to implement with
  acceptable accuracy.
• Segmentation by clustering does not require a
  training set. Therefore it is not as limited to
  image modality and application as some other
  segmentation approaches such as the active shape
  model.

5
Imaging Modalities Investigated for Automated
Segmentation by Adaptive Clustering

• Magnetic Resonance Imaging ( MRI)
• Multi-spectral Optical Imaging
• (a) Retinal Imaging
• (b) Cervicography

6
MR image segmentation
7
Background

• Computer-aided segmentation of MRI provides
  objective and quantative information about
  multiple sclerosis lesions for follow-up of
  pathology.
• Segmentation of MS lesions from a single mode
  image (T1-, T2, or PD-weighted MRI) is difficult
  because of the lack of intensity dissimilarity
  between MS lesions and other tissues. Either
  combination of all modes or masking is needed.
• Clinical MRI are affected by noise, and intensity
  inhomogeneities.

8
Retinal image segmentation
9
Background

• Stereo image pairs are captured from different
  perspective, baseline information N/A
• Depth of the retina can be extracted by
  triangulation of corresponding points in the
  stereoscopic images
• Segmentation through 3-D visualization is more
  similar to clinical diagnosis than traditional
  2-D segmentation, yielding more accurate cup/disc
  ratio

Imaging model
2-D disc/cup segmentation
3-D visualization
10
Cervical image segmentation
11
Background

• Segmentation of cervical lesion is important in
  cervical cancer detection, monitoring and drug
  administration study.
• The cervical images are taken with a regular high
  resolution color camera.
• Preprocessing to reduce or remove the glare and
  non-uniform illumination artifacts is necessary.

12
Clustering techniques

• K-means, and fuzzy C-means are some of the most
  well-known and often used clustering techniques.
  However, both are initialization dependent.
• In practical application, human interaction is
  not desirable during segmentation process.
• Adaptive fuzzy leader clustering (AFLC) and
  deterministic annealing (DA) are two approaches
  that are effective and fast with no requirement
  of initialization

13
Selected segmentation techniques

• Adaptive fuzzy leader clusteringan integrated
  neuro- fuzzy clustering technique
  (Newton,Pemmaraju,and Mitra, IEEE Trans. Neural
  Networks, 1992).
• Deterministic annealingbased on a statistical
  frame work ( K. Rose, Proc. IEEE,1998).

14
Adaptive fuzzy leader clustering

• Neural network ART-1 structure for initial
• input classification

15
Deterministic annealing
Min(F)
Clustering/optimization/cooling process when the
system reach critical temperature (TTc),
clusters split.
16
3-D Segmentation of retinal optic disc/cup

• Segmentation and visualization of optic disc/cup
  helps to quantatively detect and monitor vascular
  diseases, glaucoma hypertension, and diabetic
  retinopathy
• Better feature extraction techniques helps to
  increase the accuracy of segmentation
• Results of 3-D segmentation correlations well
  with professional manual segmentation

Cross-correlation disparity mapping
Feature extraction
Registration with power cepstrum
Window-based coarse to fine disparity computation
Visualization
17
Feature Extraction

• Blood vessels extraction is important for
  registration and disparity mapping of the stereo
  images

Old model of blood vessel segmentation using
Gaussian filtering
18
DA segmentation of retinal image
19
Blood vessel segmentation
Gaussian filtering
Segmented blood vessels
Segmented blood vessels
Feature map
Feature map
DA segmentation
Left image
Right image
20
Disparity Distribution
Disparity distribution from fundus image taken in
1994
Disparity distribution of the same patient in 1999
Change in Volume indicates progress in Glaucoma
21
Optic disc/cup segmentation
Gaussian filtering
Interpolated disparity map
disparity map
3-D visualization
DA segmentation
22
Segmentation result
Segmented disc/cup
Compared with manual segmentation
23
Validation
Data From Vertical cup length Vertical disc length Horiz. cup length Horiz. disc length Cup area Disc area Cup volume Disc volume
Patient 1 1989 198 243 153 194 22792 34779 218818 283568
Patient 1 1993 209 251 185 247 30032 46515 2904001 345188
Patient 1 1996 223 242 180 200 32110 39087 298944 331957
Patient 1 1999 226 244 176 197 31001 38174 203727 215083
Measures based on computer generated segmentation
(CG)
Patient\R R ( vertical length) (CG) R (vertical length) (MO) R (horizontal length) (CG) R (horizontal length) (MO) R (area) (CG) R (area) (MO) R (volume) (CG) R (volume) (MO)
Patient 1 1989 0.81 0.7 0.79 0.71 0.66 0.48 0.77 0.75
Patient 1 1993 0.83 0.73 0.75 0.71 0.65 0.49 0.84 0.75
Patient 1 1996 0.92 0.75 0.90 0.77 0.82 0.59 0.9 0.83
Patient 1 1999 0.93 0.75 0.89 0.78 0.81 0.58 0.95 0.93
Correlation between MO and CG 0.91 0.91 0.96 0.96 0.99 0.99 0.91 0.91
(R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated (R) Cup to disc ratio (MO)Manually Segmented by the Ophthalmologist (CG)Computer Generated
2-D and 3-D cup/disc ratios Correlation between
manual and computer segmentation
24
3D Reconstruction of the Pentagon
25
Multiple sclerosis segmentation on simulated MRI
image
T1-weighted
T2-weighted
PD-weighted
Fuzzy MS truth model
DA segmented MS
AFLC segmented MS
26
Multiple sclerosis segmentation on clinical MRI
images
MRI images in chronic order
DA segmented MS
27
AFLC DA
AFLC segmentation
classified CSF 9.01 misclassification classified gray matter 10.27 misclassification classified white matter 5.29 misclassification
DA segmentation
Noisy simulated MRI image
classified CSF 10.48 misclassification classified gray matter 10.47 misclassification classified white matter 5.10 misclassification
28
Motivation

• Color images acquired in many medical imaging are
  of poor quality for diagnostic evaluation
• A significant application of color image encoder
  is for efficient management of cervical cancer
  where visual or photographic examination, namely,
  colposcopy and cervicography are used as adjunct
  screening tests in conjunction with the Pap smear
• For optimal representation of color images, a
  histogram transformation in the V-plane, from the
  HSV conversion of the original RGB image of each
  Cervigram, is needed prior to any image analysis

29
Motivation

• Annually, there are 400,000 new cases of invasive
  cervical cancer
• 15,000 occur in the US alone.
• Cervical cancer, the second most common cancer
  affecting women worldwide and the most common in
  developing countries
• Cervical cancer can be cured in almost all
  patients, if detected by high quality repeat Pap
  screening, and treatment.

30
Motivation

• Nevertheless, cervical cancer incidence and
  mortality remain high in resource-poor regions,
  where high-quality Pap screening programs often
  cannot be maintained because of inherent
  complexity.
• Approximately 190,000 deaths per year on a
  worldwide basis
• Disease remains undetected
• Relatively insensitive, and frequently
  non-existent screening tests

31
Adjunct Screening

• Adjunct cervical cancer screening uses visual
  testing based on color change of cervix tissues
  when exposed to acetic acid.
• Cervicography has been widely used over the last
  few decades.

32
Work in Progress

• During the Guanacaste and ALTS Triage studies
  conducted by the NCI, thousands of women were
  screened by this technique and a huge amount of
  visual information have been obtained.
• During those projects, close to 100,000
  cervigrams (35 mm color pictures) were taken from
  patients with invasive cancer or intraepithelial
  lesions healthy women at enrollment who
  developed disease during the follow-up, women who
  never had pathological changes in the cervix,
  etc.

33
Work in Progress

• This information gives us a unique opportunity
  for studying the uterine cervix changes related
  or not related to HPV infection and/or cervical
  disease.
• However, each of these slides when digitized,
  requiring 11 MB to store. A Web-based digital
  archive for sharing and evaluating these
  high-quality pictures must be designed, since the
  large sizes make them impractical for efficient
  storage and processing at a single site.

34
Work in Progress

• A non-profit alliance among medical experts from
  the NCI, the Medical College of Georgia, digital
  archive experts from the National Library of
  Medicine and experts from TTU, well-known for
  their work on digital image compression, has been
  formed to address the optimum management and
  early detection of cervical cancer world-wide
  utilizing this cervix data set.

35
PSNR Vs Compression Levels
36
Cervical lesion segmentation
Manual segmentation
AFLC segmentation
DA segmentation
37
Conclusions

• Clustering techniques investigated can be
  effectively applied to segmentation of medical
  images from different modalities
• Good correlation between computer segmentation
  and manual segmentation suggests promising
  application in computer-aided diagnosis.

38
References

• Clarke, L. P., Camacho, R. P., Velthuizen, M. A.,
  Heine, J. J., Vaidyanathan, M., Hall, L. O.,
  Thatcher, R. W. and Silbiger, M. L., Review of
  MRI segmentation methods and applications,
  Magn. Reso. Imaging, Vol. 13, No. 3, pp. 343-368,
  1995.
• Pal N. R. and Pal, S. K., A review on image
  segmentation techniques, Pattern Recognition,
  Vol. 26, No. 9, pp. 1277-1294, 1993.
• Haralick, R.M. and Shapiro, L. G., Image
  segmentation techniques, Computer Vision,
  Graphics, and Image Processing, Vol. 29, No. 1,
  pp. 100-132, Jan, 1985.
• L. A. Zadeh, Outline of a new approach to the
  analysis of complex systems and decision
  processes, IEEE Trans. On Syst. ,Man, Cybern.,
  Vol. SMC-3, PP 28-44,1973.
• Bezdek, J.C., Hall, L.O. and Clarke, L.P. Review
  of MR image segmentation techniques using pattern
  recognition, Med. Phys., Vol. 20, pp.
  1033-1048, 1993.
• Carpenter, G. A. and Grossberg, S., A massively
  parallel architecture for a self-organizing
  neural pattern recognition machine, Computer
  Vision, Graphics and Image Processing, Vol. 37,
  pp. 54-115, 1987.
• Newton, S. C., Pemmaraju, S. and Mitra, S.,
  Adaptive fuzzy leader clustering of complex data
  sets in pattern recognition, IEEE Trans. Neural
  Networks , Vol. 3, pp.794-800, 1992.
• Mitra, S., Castellanos, R., Yang, S.Y. and
  Pemmaraju, S., Adaptive clustering for image
  segmentation and vector quantization, book
  chapter in Soft-computing for Image Processing,
  Editors Pal, S. K., Ghosh, A. and Kundu, M. K.
  New York Springer-Verlag, 1999.
• Rose, K. Deterministic annealing for clustering,
  compression, classification, regression, and
  related optimization problems, Proceedings of the
  IEEE, Vol. 86, No. 11, Nov. 1998.
• Bezdek, J., Pattern Recognition with Fuzzy
  Objective Function Algorithms, New York Plenum
  Press, 1981.
• Enrique Corona, Sunanda Mitra, Mark Wilson,
  Thomas Krile, Young H. Kwon, Peter Soliz,
  Digital Stereo Image Analyzer for Generating
  Automated 3-D measures of Optic Disc Deformation
  in Glaucoma, IEEE Transactions on Medical
  Imaging, Special Issue on Image Processing and
  Analysis in Drug Discovery and Clinical Trials,
  Vol.21, No.10, pp1244-1253, October, 2002.
• Shuyu Yang and Sunanda Mitra, Content Based
  Vector Coder for Efficient Information
  Retrieval, Invited Book Chapter in New
  Directions in Enhancing the power of the
  Internet, EditorsMasoud Nikravesh, B. Azvine,
  R. Yagar, and Lotfi A. Zadeh. Physica-Verlag,
  Springer, November, 2003.
• Shuyu Yang, Sunanda Mitra, Enrique Corona, Brian
  Nutter, and D. J. Lee,"Multi-level Wavelet
  Feature Statistics for Efficient Retrieval,
  Transmission, and Display of Medical Images by
  Hybrid Encoding", EURASIP Journal on Applied
  Signal Processing,Vol.5,pp 449-460, 2003.
• Cocosco, C.A., Kollokian, V., Kwan, R.K.-S. and
  A.C. Evans, BrainWeb Online Interface to a 3D
  MRI Simulated Brain Database, http//www.bic.mni.
  mcgill.ca/brainweb/
• Johnson, K.A. and Becker, J. A. The whole brain
  atlas, http//www.med.harvard.edu/AANLIB/.