A New Image-Mining Technique for Automation

of Parkinson’s Disease Research

Contribution
1 We proposed new technique and created on its base
the algorithmic scheme for automated extraction of data
for construction and investigation of Parkinson’s disease
experimental model.
2 Experimental investigations of the scheme confirmed:
possibility and suitability of striatum frontal section images
automated processing and analysis;
the precision of the results is comparable (non less than)
Igor Gurevich Artem Myagkov Vera Yashina
Dorodnicyn Computing Centre of the Russian Academy of Sciences,
Moscow, The Russian Federation
igourevi@ccas.ru, artem.myagkov@gmail.com, werayashina@gmail.com
The Proposed Algorithmic Scheme Advantages of Image Processing and Analysis
Steps of the Algorithmic Scheme
The developed algorithmic scheme consists of the following stages:
1 preprocessing: 1.1) opening by reconstruction; 1.2) the bot-hat transformation by dual reconstruction; Preprocessing and image analysis stages of the scheme
1.3) closing by dual reconstruction; 1.4) h-dome elimination transformation; are based on the mathematical morphology theory.
2 image analysis: 2.1) object and background markers extraction; 2.2) morphological gradient image
High efficiency of the developed technique is caused by:
modification; 2.3) watershed segmentation; 1 computational efficiency of implementation of
3 construction of object representations: 3.1) construction of feature descriptions; 3.2) feature selection; morphological operators;
4 classification: 4.1) objects clustering; 4.2) results interpretation. 2 selected sequence (combination) of transformations;

with that of manual object feature estimation. 3 the ordering of used structuring elements.

3 Experimental applications of the developed technique Application of this stages of the scheme to initial
Preprocessing — Explanation
confirmed its high efficiency: images enables:
150 300 260 260

200 times increase in productivity and 10 times decrease 140

250
250 250
to smooth complex background;
240

in the amount of experimental animals and expendables. Gray value 130 240
230
to detect small objects on images;

Gray value

Gray value

Gray value
200 Opening by reconstruction
120 230
dual reconstruction 220
Bot−hat by dual reconstruction
110 220

100
Initial image
erosion
150

210
Bot−hat by dual reconstruction
dilation
Closing by dual reconstruction
210

200 Closing by dual reconstruction

to eliminate out-of-focus objects;
Opening by reconstruction
100

Initial Data
90

80 50
200

190
190

180
h subtraction, h = 18
H−max elimination to separate close objects.
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Y Y Y Y

The initial data is digital Experimental data have been Step 1.1 — Elimination of Step 1.2 — Correction of Step 1.3 — Nonuniform Step 1.4 — H-dome
images of sections of various received from digital images initial image narrow peaks complex heterogeneous regions smoothing in the elimination corresponds to Visual Comparison of Automatically and
brain areas. of distal parts of axons. corresponding to the background on the initial interior of the terminals out-of-focus objects Manually Selected Objects
background image removal
I1 = ρiI(Ii B1) I2 = ρ∗I1(I1 ⊕ B2) − I1 I3 = ρ∗I2(I2 ⊕ B3) I4 = I3 − ρI3(I3 − h)
B1 = disk (0.5(r + R)) B2 = disk(R) B3 = disk(r) h = tresh ({min(I3)})

Accuracy Estimation
Statistics Object extraction method Area (µm2) Mean intensity Number of Terminals
Neurons Terminals Experiment Control Experiment Control Experiment Control
2 2
1 The resolution of initial images is 0.0117 µm /pixel .
Mean Automated 1.37 1.30 96.23 82.93 15.2 36.4
2 Terminals are rounded objects with area varying from
Manual 1.56 1.52 95.79 83.60 13.6 34.6
0.6 − 0.7 µm2 up to 2.5 − 3 µm2 .
Standard Automated 0.69 0.52 9.93 10.77 3.1 3.8 Automated terminal Manual terminal extraction
deviation Manual 0.57 0.70 10.99 11.57 3.8 4.0 extraction
Automation of Parkinson’s Disease Model Null Hypothesis rejected rejected accepted accepted accepted accepted
Research Acknowledgments
Construction of Object Representations and Classification
The problem —
the development and study of suitable mathematical technique This work was partially supported by the Russian
for automated extraction and characterization of objects in brain 25 morphometric, densitometric and textural features terminals were clustered by construction a committee Foundation for Basic Research Grant No. 09-07-13595,
section images; were used of different modern clustering techniques: hierarchical by the Program of Fundamental Research
the implementation of the technique as an efficient algorithmic a successive algorithm for feature selection was used clustering algorithms, mixture-resolving algorithms, of the Presidium of the Russian Academy of Sciences
scheme and software.
to find the most discriminate features when nearest neighbor clustering, evolutionary approaches (2010) and by the program “Participant of Young
The approach — joint use of:
the terminals are classified into experiment and control for clustering and search-based approaches. Scientific-Innovate Tender” (“U.M.N.I.K.”)
image processing and analysis techniques of mathematical
morphology; groups 2 and 5 well-separated groups were selected of the Foundation for Assistance to Small Innovative
mathematical theory of pattern recognition; the accuracy of classification was increased from 67% Enterprises (state contract Nos. 6956p-9009,
Descriptive Approach to image analysis. to 74%. 6957p-9010).

Vavilov st. 40, 119333 Moscow, Russian Federation — DEPARTMENT of MATHEMATICAL and APPLIED METHODS of IMAGE ANALYSIS Mail: igourevi@ccas.ru WWW: http://www.ccas.ru
and NONLINEAR PROBLEMS — Institution of Russian Academy of Sciences Dorodnicyn Computing Centre of RAS