Using black and white models for classification of medical images

1
Using black and white models for classification
of medical images

• Sergei Kucheryavski, Altai State University,
  Russia
• svk_at_asu.ru

2
Prehistory analysis of medical data

• Children's hospital of Altai region
• analysis of frequencies of different diseases
  occurring in patients with perinatal lesions of
  the central nervous system
• analysis and recognition of blood cells
• analysis and recognition of marrow cells
• Department of urology and nephrology of AMU
• analysis of ultrasound and X-ray tomograms of
  urolitas
• determination of operable/therapeutic state of
  disease using ultrasound and X-ray tomograms of
  tissues

3
Prehistory existent methods of analysis

• Based on the conceptual models
• geometrical size of objects and distance between
  them
• geometrical area of objects or segments
• hue and color intensity
• Rigid requirements to the raw data
• low-noise
• high level of contrast and intensity

Morphology a a branch of biology that deals
with the form and structure of animals and
plants b the form and structure of an organism
or any of its parts Merriam-Webster Online
Dictionary
4
Presummary and questions to answer

• Habitual methods, based on the hard model of
  studied objects, are very spread in medicine
• Soft models based image analysis approach usually
  allows to analyze images with middle and low
  quality including noised images
• Therefore
• Is it possible to use soft model approach for
  medicine purposes?
• Will such approach give results with acceptable
  quality?
• Are there any advantage in using soft model
  approach in comparison with traditional one?

5
Blood cells formation
red cells
white cells
6
Basic white cell types
segmented
row
Neutrophils
Lymphocyte
Monocyte
7
Morphology analysis

• cell area
• kernel area
• cell hue
• kernel hue
• skeleton radius
• kernel min thickness
• kernel max thickness
• number of kernel segments

raw image
segmentation
edge/skeleton detection
properties
8
Blood cells analysis software

• Conditions
• rigid requirements to the image quality
• sensitive to presence of noise
• rigid requirements to the smear quality
• Effects
• poor results for middle and low quality images
• rigid requirements to the equipment (microscopes,
  cameras, etc)
• rigid requirements to chemical for smear
  preparation
• As a result
• highly recommended to use such software with
  equipment and chemical from the same producer
• price for software only 2 000 10 000
• price for equipment 50 000 100 000

9
Ordinary and good pictures of blood
10
Classification algorithm
acquisition
preprocessing
features extracting
classification

• Digital cameras
• Video capturing and digitizing

• Segmentation
• Contrast stretching
• Brightness enhancement

• Wavelet transformation
• AMT

• PCA
• PLS-DA

11
Features vector building

• Wavelet transformation
• transforms image from spatial to
  frequency-spatial domain
• good results in different areas of image
  recognition and analysis
• quick and simple algorithm
• AMT
• transforms image from spatial to scale domain
• good result in classification of both
  heterogeneous images and textures
• simple algorithm but relatively slow for big
  images (1-4 seconds in comparison with Wavelet
  transformation - 0.2-0.8 seconds)

12
Features vector wavelet transformation
H gives smoothing signal G gives the details
1D signal
2D signal
13
Features vector wavelet transformation

• For feature vector we calculate metrics of
  horizontal, vertical and diagonal details
• Feature vector f(dh1),f(dv1),
  f(dd1),,f(dhm),f(dvm),f(ddm)
• 1m level of wavelet transform
• dh, dv, dd horizontal, vertical and diagonal
  details
• f() metrics function
• Useful metrics
• Energy
• Standard deviation
• Moments

14
Features vector AMT

• Was developed by Robert Andrle as a substitute of
  fractal analysis for the purpose of complexity of
  geomorphic lines investigation (R. Andrle, Math.
  Geol., 16, 83-79, (1996))
• Was introduced into chemometrics as generic
  approach for analysis of measurement series by
  Esbensen et al(K.H. Esbensen, K, Kvaal, K.H.
  Hjelmen, J. Chemom., 10, 569-590, (1996))
• Properties
• transforms the 2D image into 1D spectra without
  losses the structure information
• highly sensitive for changing of typical scales
  of objects on images

15
Features vector AMT
Step 1 Unfolding
16
Features vector AMT
Step 2 Sampling
Step 3 Measure angle and calculation mean angle
for all points
17
Features vector AMT

• Step 4 Change radius S and repeat step 3 for
  mean angle vector (spectrum) building
• Mean angle values (MAS) for each S from S0 to SM
  compose mean angle spectrum MAS0,,MASM
• Example of MA spectrum
• Spectrum can be regarded as a vector of images
  features on set of scales

18
Objects for investigations

• Calibration set
• 60 samples
• 2 classes
• Samples were taken from different people
• Ordinary microscope and cheap VGA camera were
  used
• Test set
• 96 samples
• Samples were taken from different people
• Samples were taken in other day then calibration
  set
• Ordinary microscope and cheap VGA camera were
  used

19
Objects for investigations
20
Preliminary PCA and PLS (calibration set)
AMT
Wavelet transform
21
Preliminary PLS (test set)
AMT
Wavelet transform
22
Surf of blood cells
23
Spiral unfolding
24
Unfolded images profiles
25
Preliminary analysis
26
PLS-DA results

• Prediction of calibration set
• 60 samples
• Samples were taken in different days and from
  different people

• Prediction of test set
• 96 samples
• Samples were taken in different days and from
  different people

27
Summary

• Conclusions
• Hard-modeling approach that is used to image
  analysis effective only for high-quality images
• The soft-modeling approach of image
  classification was applied to the task of blood
  cell type recognition on low-quality images
• The effectiveness of recognition was 96-97 that
  allows to speak about advantages of such approach
• To be continued
• Analysis of middle resolution images (1-2 Mp)
• Approximation of cells by ellipse curve and
  ellipse-like unfolding
• Use other methods for analysis of image profiles

28
Acknowledgements

• Alexey Pijanzin, docent, doctor of Children's
  hospital of Altai region
• Ivan Belyaev, M.S. student of Altai State
  University
• Sergei Zhilin, PhD, docent of Altai State
  University