Rayleigh Mixture Model and its Application for Ultrasound-based Plaque Characterization

1
Rayleigh Mixture Model and its Application for
Ultrasound-based Plaque Characterization
Workshop Programa Doutoral em Engenharia
Biomédica 15 Julho 2009

• José Seabra, Francesco Ciompi, Oriol Pujol, Petia
  Radeva and João Sanches
• Instituto de Sistemas e Robótica, IST Lisboa
• Centre de Visió per Computador, Barcelona

2
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• Vulnerable plaques are a major source of carotid
  and coronary circulatory events
• B-mode ultrasound and IVUS provide accurate
  representation of the arterial wall and plaque
• Identification of vulnerable plaques comes from
  the correct modeling of tissue echo- morphology
  and characterization of its composition
• Under particular conditions, pixel intensity
  observations belonging to ultrasound images are
  well modeled by Rayleigh probability density
  functions (pdfs)
• Proposal
• to characterize the echo-morphology of plaques by
  use of a mixture of Rayleigh distributions
• to incorporate the Rayleigh Mixture Model (RMM)
  in a 3-type plaque classification problem

3
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• A plaque (as other tissue) can be regarded as a
  complex structure (see Fig.1) where phenomena,
  including absorption, diffuse and structural
  scattering, occur and combine

Figure 1. Tissue Acoustic model
4
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

Figure 2. Effect of the Rayleigh reflectivity
parameter on the pdf
5
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• Simulation study for testing the RMM in a
  synthetic image

Figure 3. a) Tissue sample and b) diagonal D
intensity profile. c) MLE of the Rayleigh pdf for
region S, and d) comparison between MLE Rayleigh
pdf and mixture pdf for the whole tissue sample
6
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• Plaque characterization is based on an IVUS study
  of the coronary arteries
• Features are based on images reconstructed from
  the RF data
• 67 plaques were labeled according to their
  composition as lipidic, fibrotic or calcified

(d)
(e)
Figure 4. a) IVUS data acquisition and analysis
from a post-mortem human coronary artery. B)
Histological analysis of a slice of the artery.
(c) a reliable correspondence in the IVUS image
is established by using a suitable labeling
software. (d) Rotation catheter, (e) Polar vs
reconstructed IVUS image
7
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• RMM estimation for 3 different plaques,
  generation of de-speckled and speckle images

(a) (b) (c)
(d) (e) (f)
Figure 5. a) IVUS image showing three plaques
(tissues) labeled according to their composition.
(b-c) De-speckle and speckle image (the
regularization effect is visible). (d-e) RMM
estimated from the three labeled distinct plaques
8
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• Performance was evaluated by means of the
  Leave-One-Patient-Out (LOPO) cross-validation
  technique, using the Adaboost classifier with
  Error-Correcting-Output Codes (ECOC)
• 1st Result Plaque-content classification using
  RMM features

(b)
(a)
Figure 6. a) Feature space, where the dataset of
67 plaques of different types is plotted
according to the mixture coefficients. b) 3-type
plaque-content characterization using RMM
computed with different number of mixture
components
9
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• 2nd Result Local-wise classification by use of a
  feature set including RMM, Speckle, Textural and
  Spectral features

(a)
Features Weights
RMM Speckle Texture
Spectrum
(b)
(c)
Figure 7. a) Feature space, where the dataset of
67 plaques of different types is plotted
according to the mixture coefficients. b) 3-type
plaque-content characterization using RMM
computed with different number of mixture
components. c) Graphical classification
10
Introduction Rayleigh Mixture Model Plaque Classification Results Conclusions

• A generic method to model the tissue
  echo-morphology is proposed based on the mixture
  of Rayleigh distributions
• Our study suggests that different plaque types
  can be distinguished based on the coefficients
  (weights) and Rayleigh parameters of each
  distribution of the mixture
• The inclusion of mixture parameters in a
  classification framework has shown to improve the
  discriminative power between different plaque
  types, leading to high classification accuracies
• A medical supervised plaque classification tool
  based on RMM can be developed, given that what is
  suspected to be a plaque is previously segmented
  and provided to the algorithm
• FUTURE
• Change from Rayleigh mixture to Rician mixture
• Apply this mixture concept and its features to
  classification of symptomatic carotid plaques