49
A Classification Method for TCA-Images München September 15, 2005 - p. 1/15 A Classification Method for TCA-Images 6. Kongress der Gesellschaft für Anthropologie e.V. "Facetten der modernen Anthropologie" Katy Streso Max Planck Institute for Demographic Research www.demogr.mpg.de

A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

A Classification Method for TCA-Images München September 15, 2005 - p. 1/15

A Classification Method for TCA-Images6. Kongress der Gesellschaft für Anthropologie e.V.

"Facetten der modernen Anthropologie"

Katy StresoMax Planck Institute for Demographic Research

www.demogr.mpg.de

Page 2: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 2/15

Introduction

■ Introduction to Tooth Cementum Annulation (TCA) Methodand Images

■ The Statistical Model - HMRF■ Application

Page 3: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 2/15

Introduction

■ Introduction to Tooth Cementum Annulation (TCA) Methodand Images

■ The Statistical Model - HMRF

■ Application

Page 4: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 2/15

Introduction

■ Introduction to Tooth Cementum Annulation (TCA) Methodand Images

■ The Statistical Model - HMRF■ Application

Page 5: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 3/15

Introduction to Tooth CementumAnnulation (TCA) Method and Images

Page 6: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

- --

■ age estimation method■ paleodemographers: want to reconstruct mortality profiles of

historical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 7: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

-

--

■ age estimation method■ paleodemographers: want to reconstruct mortality profiles of

historical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 8: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

- -

-

■ age estimation method

■ paleodemographers: want to reconstruct mortality profiles ofhistorical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 9: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

- --

■ age estimation method■ paleodemographers: want to reconstruct mortality profiles of

historical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 10: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

- --

■ age estimation method■ paleodemographers: want to reconstruct mortality profiles of

historical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 11: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 4/15

TCA Method and Images

- --

■ age estimation method■ paleodemographers: want to reconstruct mortality profiles of

historical populations by counting tooth rings→ need of an objective and automatic evaluation (Diss)

■ TCAa images◆ ≈ 1016 x 1300 pixels◆ gray values [0, 28 − 1] or [0, 212 − 1]◆ tooth ring roughly 1-3 µm (5-20 pixel) thick

■ typical good quality TCA-image:

a[Hoppa and Vaupel, 2002]

Page 12: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 5/15

TCA Image

Page 13: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 6/15

TCA-Image Analysis

■ typical result after thresholding:

need to punish saw cuts, reinforce tooth rings

■ incorporate information about neighboring pixel◆ Fourier transformer

(applied on TCA-images by Czermak a)◆ set up statistical model to include spatial dependencies

aCzermak, A. (2004). Automatisierte Auszahlung von Zahnzementzuwachsringen

(TCA). Talk presented at the Appa-Tagung 2004 but not yet published. See

http://www.gfanet.de/docs/appa workshop 10 04 beitraege.pdf.

Page 14: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

● TCA Method and Images

● TCA Image

● TCA-Image Analysis

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 6/15

TCA-Image Analysis

■ typical result after thresholding:

need to punish saw cuts, reinforce tooth rings■ incorporate information about neighboring pixel

◆ Fourier transformer(applied on TCA-images by Czermak a)

◆ set up statistical model to include spatial dependenciesaCzermak, A. (2004). Automatisierte Auszahlung von Zahnzementzuwachsringen

(TCA). Talk presented at the Appa-Tagung 2004 but not yet published. See

http://www.gfanet.de/docs/appa workshop 10 04 beitraege.pdf.

Page 15: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 7/15

The Statistical Model - HMRF

Page 16: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 17: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

true, unknown label image

TCA-imageIobs

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 18: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 19: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 20: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue)

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 21: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 22: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 23: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 24: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 25: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

unknown label imageItrue

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

= µ

+

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 26: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

=

µ

+unknown label imageItrue

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

µ

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 27: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

=

µ

+unknown label imageItrue

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

µ

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)

a[Zhang et al., 2001]

Page 28: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 8/15

Hidden Markov Random Field (HMRF) Model a

TCA-imageIobs

0 0 1 0 1 1 0

P (Itrue) ∼ MRF

´6contextualconstraints

independent noiseInoise

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

=

µ

+unknown label imageItrue

P (Inoise) ∼∏

(x,y)

N(

0, σ2)

µ

²±¯°

■ maximize posterior distribution (computationally expensive)

P (Itrue|Iobs) ∝ P (Itrue)P (Inoise)

■ specify MRF ! (include prior knowledge about tooth rings)a[Zhang et al., 2001]

Page 29: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)◆ use a bank of filters

with variable ring width T◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 30: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)◆ use a bank of filters

with variable ring width T◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 31: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)◆ use a bank of filters

with variable ring width T◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 32: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)

◆ use a bank of filterswith variable ring width T

◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 33: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)◆ use a bank of filters

with variable ring width T

◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 34: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 9/15

Markov Random Field (MRF) Model

■ Markov-property: Itrue

X

■ FRAMEa

◆ special kind of MRF◆ Filters, Random Fields and Maximum Entropy◆ uses filters to collect neighborhood information

■ Filter◆ at each pixel: measure similarity of neighborhood to filter

(by convolution)◆ use a bank of filters

with variable ring width T◆ select best T during maximization

a[Zhu and Mumford, 1997],[Zhu et al., 1997],[Zhu et al., 1998]

Page 35: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 10/15

Markov Random Field (MRF) Model - FRAME

■ prior distribution (incorporates prior knowledge in filter F )

P (Itrue) =1

Ze∑

(x,y) |(F∗Itrue)(x,y)|

■ typical prior assumption about TCA-image (Gibbs simulation)

Page 36: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

● Hidden Markov Random Field

(HMRF) Model

● Markov Random Field (MRF)

Model● Markov Random Field (MRF)

Model - FRAME

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 10/15

Markov Random Field (MRF) Model - FRAME

■ prior distribution (incorporates prior knowledge in filter F )

P (Itrue) =1

Ze∑

(x,y) |(F∗Itrue)(x,y)|

■ typical prior assumption about TCA-image (Gibbs simulation)

Page 37: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 11/15

Application

Page 38: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 39: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 40: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 41: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings

■ bifurcations:where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 42: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 43: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 44: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 12/15

Application

■ theor. ] rings: 38recognized: ≈ 35

■ miss thin rings■ bifurcations:

where tooth ringshave differentorientation

■ → reconstructionheavily influencedby filter F

■ global property →select locationdependent filters

Page 45: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

● Application

● Bibliography

A Classification Method for TCA-Images München September 15, 2005 - p. 13/15

Bibliography

[Hoppa and Vaupel, 2002] Hoppa, R. D. and Vaupel, J. W., editors (2002).Paleodemography: Age Distributions from Skeletal Samples. CambridgeUniversity Press, Cambridge.

[Zhang et al., 2001] Zhang, Y., Brady, M., and Smith, S. (2001). Segmenta-tion of Brain MR Images Through a Hidden Markov Random Field Modeland the Expectation-Maximization Algorithm. IEEE Transactions on Med-ical Imaging, 20(1):45–57.

[Zhu and Mumford, 1997] Zhu, S. C. and Mumford, D. B. (1997). PriorLearning and Gibbs Reaction-Diffusion. IEEE Transactions on PatternAnalysis and Machine Intelligence, 19(11):1236–1250.

[Zhu et al., 1998] Zhu, S. C., Wu, Y., and Mumford, D. B. (1998). Fil-ters, Random Fields and Maximum Entropy (FRAME): Towards a Uni£edTheory for Texture Modeling. International Journal of Computer VisionArchive, 27(2):107 – 126.

[Zhu et al., 1997] Zhu, S. C., Wu, Y. N., and Mumford, D. B. (1997). Min-imax Entropy Principle and Its Application to Texture Modeling. NeuralComputation, 9(8):1627–1660.

Page 46: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

A Classification Method for TCA-Images München September 15, 2005 - p. 14/15

Page 47: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

Fourier Transform

● Fourier Transform

A Classification Method for TCA-Images München September 15, 2005 - p. 15/15

Fourier Transform

■ is a directional and global methodThis could introduce a substantial error, if◆ direction of tooth rings is not horizontal (or vertical)◆ rings are changing directions too heavy across the image◆ there exists directional noise, not orthogonal to tooth rings

(noise can not be removed without over-smoothing)◆ image contains thin and thick rings (removing a frequency

removes rings of a certain size from the whole image)It is herewith also erroneous to smooth a whole TCAimage, including parts where no rings exist like the dentin.

■ Fourier frequencies can not be translated to a person’s age.Because of superposition of sine and cosine waves we cannot directly interpret one frequency of the Fourier transforminto one ring width or one ring count. Cutting out certainfrequencies therefore does not have an explicit meaning forTCA image analysis.

■ back to the presentation

Page 48: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

Fourier Transform

● Fourier Transform

A Classification Method for TCA-Images München September 15, 2005 - p. 15/15

Fourier Transform

■ is a directional and global methodThis could introduce a substantial error, if◆ direction of tooth rings is not horizontal (or vertical)◆ rings are changing directions too heavy across the image◆ there exists directional noise, not orthogonal to tooth rings

(noise can not be removed without over-smoothing)◆ image contains thin and thick rings (removing a frequency

removes rings of a certain size from the whole image)It is herewith also erroneous to smooth a whole TCAimage, including parts where no rings exist like the dentin.

■ Fourier frequencies can not be translated to a person’s age.Because of superposition of sine and cosine waves we cannot directly interpret one frequency of the Fourier transforminto one ring width or one ring count. Cutting out certainfrequencies therefore does not have an explicit meaning forTCA image analysis.

■ back to the presentation

Page 49: A Classification Method for TCA-Images · Introduction Introduction to Tooth Cementum Annulation (TCA) Method and Images The Statistical Model - HMRF Application A Classification

● Introduction

Introduction to Tooth Cementum

Annulation (TCA) Method and

Images

The Statistical Model - HMRF

Application

Fourier Transform

● Fourier Transform

A Classification Method for TCA-Images München September 15, 2005 - p. 15/15

Fourier Transform

■ is a directional and global methodThis could introduce a substantial error, if◆ direction of tooth rings is not horizontal (or vertical)◆ rings are changing directions too heavy across the image◆ there exists directional noise, not orthogonal to tooth rings

(noise can not be removed without over-smoothing)◆ image contains thin and thick rings (removing a frequency

removes rings of a certain size from the whole image)It is herewith also erroneous to smooth a whole TCAimage, including parts where no rings exist like the dentin.

■ Fourier frequencies can not be translated to a person’s age.Because of superposition of sine and cosine waves we cannot directly interpret one frequency of the Fourier transforminto one ring width or one ring count. Cutting out certainfrequencies therefore does not have an explicit meaning forTCA image analysis.

■ back to the presentation