estimation of age in forensic medicine using multivariate data analysis ivan belyaev altai state...

Post on 05-Jan-2016

216 Views

Category:

Documents

1 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Estimation of age in forensic medicine using multivariate data

analysis

Ivan Belyaev

Altai State University, Barnaul, Russia

iab@asu.ru

Sergey Kucheryavsky

Aalborg University Esbjerg, ACABS group, Esbjerg, Denmark

WSC'6Kazan–2008

Outline

Introduction Approach which used for age and sex

identification in forensic medicine Description of proposed solution, using

multivariate data analysis Results and discussion

2

2

Introduction

Skeletons' bones are resistant to environment transformations in course of time

Osteoporosis is an essential factor for age-specific changes, characterized by decrease of bone tissue density

3

Osteoporosis marks

Sponge

Cortical

Bone structure

4

Age

Sp

on

ge

den

sity

22 years

56 years

83 years

Photocolorimeter scheme

PhotodetectorLight sourceSample

6

Image classification procedure

imageacquisition

imagepreprocessing features extracting classification

Digitalcamera,

Scaner

SegmentationWavelet transform,

AMTPLS

7

Image acquisition and preprocessing

Digital camera Scanner

8

Building features vector

Sponge part of bone has textural structure Therefore, the following was chosen:

Feature vector, based on wavelet transformation good results in different areas of image recognition and analysis quick and simple algorithm

Feature vector, based on AMT 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)

9

Features vector: wavelet transformation

Raw signal

Smoothed Details

H G

Smoothed Details

H G

H – gives smoothed signalG – gives the details

1D signal

Gr

Hr

diag

hor

ver

HrHc HrGc GrHc GrGc

r — rows

c — columns

2D signal

10

Example of image WT

WT

11

Features vector: wavelet transformation

For feature vector we calculate metrics ofhorizontal, vertical and diagonal details:

Feature vector = [ f(dh1),f(dv

1), f(dd

1),…,f(dh

m),f(dv

m),f(dd

m) ]

1…m level of wavelet transformdh, dv, dd horizontal, vertical and diagonal detailsf() metrics function

Useful metrics: Energy Standard deviation Entropy

12

Experiments Samples:

Sponge tissue of I-V lumbar vertebra bodies without evident bone pathology

Samples were taken from 70 males and 62 females Aged from 21 to 93

Models:

Name Response

A-top AMT top age PLSA-bott AMT bottom age PLSA-both AMT both age PLSW-top WT top age PLSW-bott WT bottom age PLSW-both WT both age PLS

Features extraction

method

Lamina sides

Model constructing

method

13

Typical samples

Male,26 years

Male,40

years

Male,57

years

Male,71

years

A-b

oth

mod

el,

fem

ale

set

First resultsW

-bot

h m

odel

, fe

mal

e se

t

The basic reason of poor results

In our experiments the response variable is chronological age

Pores transformations reflect genuine grade of organism ageing, which in medicine is described as biological age

Biological age ≠ Chronological age

16

Correlation0.990.98

Correlation0.990.98

W-b

oth

mod

el,

fem

ale

set

A-b

oth

mod

el,

fem

ale

set

17

A-b

oth

A-t

op

A-b

ott

W-b

oth

W-t

op

W-b

ott0%

25%

50%

75%

100%

Less than 1 from 1 to 2 from 2 to 3

From 3 to 4 More than 5

A-b

oth

A-t

op

A-b

ott

W-b

oth

W-t

op

W-b

ott

Ph

0%

25%

50%

75%

100%

less than 5 from 5 to10 from 10 to 20 more than 20

Statistics for male samples prediction

Calibration set Full set

18

Statistics for female samples prediction

Full setCalibration set

A-b

oth

A-t

op

A-b

ott

W-b

oth

W-t

op

W-b

ott

Ph

0%

25%

50%

75%

100%

less than 5 from 5 to10 from 10 to 20 more than 20

A-b

oth

A-t

op

A-b

ott

W-b

oth

W-t

op

W-b

ott0%

25%

50%

75%

100%

Less than 1 from 1 to 2 from 2 to 3

From 3 to 4 More than 519

Problems and possible solutions

Problems

Not enough samples – can’t use test validation

Poor representativenes

Possible solution

use additional features

use other methods of analysis

20

top related