Transcript ppt - Simon Fraser University

SPIE
Medical Imaging
2010
Graph-based Pigment Network Detection
in Skin Images
Maryam Sadeghi1,3, Majid Razmara1, Martin Ester1,
Tim K. Lee1,2,3 and M. Stella Atkins1
1: School of Computing Science, Simon Fraser University
2: Department of Dermatology and Skin Science, University of British Columbia
3:Cancer Control Research, BC Cancer Agency
1
Skin cancer and melanoma
 Skin cancer : most common of all cancers
 Melanoma : leading cause of mortality
 Early detection: significantly reduces mortality
Basal cell carcinoma
2
Combined nevus
[ Images courtesy of “Dermoscopy of pigmented skin lesions” ]
Melanoma
Dermoscopy
3
Pigment Network Detection
 Present (Typical or Atypical Pigment Network )
 Typical: “light to dark-brown network with small, uniformly spaced network holes and thin network lines
distributed more or less regularly throughout the lesion and usually thinning out at the periphery”
 Atypical: “black, brown or gray network with irregular holes and thick lines“
 Absent: There is no typical or atypical pigment network
4
Present
5
Absent
6
Problem Statement and Motivation
 Problem:
 Pigment network detection in dermoscopy images
 Motivation:
 Skin texture analysis for computer-aided
diagnosis
 Pigment Network Visualisation for Training
purposes
7
Algorithm overview
 Given a dermoscopy image
Original
8
Algorithm overview
 Pre-processing. Using LoG sharp changes of colors are
detected
Original
9
Laplacian of Gaussian
Algorithm overview
 Converting the result of the pre-processing to a graph.
Original
10
Laplacian of Gaussian
Image to Graph
Algorithm overview
 Converting the result of the pre-processing to a graph.
Original
11
Laplacian of Gaussian
Image to Graph
Cyclic Subgraphs
Algorithm overview
 Converting the result of the pre-processing to a graph.
Original
12
Laplacian of Gaussian
Pigment Network
Image to Graph
Cyclic Subgraphs
Algorithm overview
 Converting the result of the pre-processing to a graph.
Original
Laplacian of Gaussian
Image to Graph
Present
13
Classification
Pigment Network
Cyclic Subgraphs
Given Image
14
Filtered by Laplacian of Gaussian
15
Binary Image to Graph Conversion
 A binary image has some connected components
 Each of them is converted to a graph (G)
 Each pixel a node of G
 A unique label according to its coordinate
 Graph Gi
 |V|= size of the connected component i in pixels
7x7
 |E|=Number of edges connecting
the white pixels
|V|=17
|E|=17
3
8
 Iterative Loop Counting Algorithm.
Graph Gi
48
16
Connected Component i
Removing Undesired Cycles
 Labels of nodes
coordinates in the image
 Mean intensity of meshes in the original image
 Globules and dots: Inside color darker than outside color
 Inside Color
 Outside Color
 Extended area by 2 pixels
 Tuning the Thresholds
17
18
19
Pigment Network Graph
 A new graph representing the Pigment Network
 Centers of the detected cycles ( green holes in the
image) are determined as nodes
 For each center the distance from all nodes is computed
 According to the size of the lesion and the average size
of the net holes, Maximum Distance Threshold (MDT)
is set
 Two nodes are connected together if they are closer than
20
MDT
21
Image Classification
 Density Ratio of the detected pigment network
Density 
E
V * log( LesionSize )
 Lesion Size: Size of the area of the image that is
inspected for finding the pigment network
 Density Threshold
 Density Ratio ≥ Threshold => Present
22
 Density Ratio < Threshold => Absent
Experimental Results
Original Image
LOG Edge Detector
Cyclic Subgraphs
Present
23 Original Image
LOG Edge Detector
Cyclic Subgraphs
Absent
Evaluation Data Set and Result:
 A set of 100 dermoscopic images used for tuning the
parameters and thresholds of the method
 500 images of size 768x512 are used to test the
performance of the method
 Taken from Argenziano et al.’s Interactive Atlas of Dermoscopy
 Each image is labeled as ‘Absent’ or ‘Present {typical, atypical}
 Accuracy: 92.6%
24
Future Work
 Features of pigment networks
 Color, regularity, thickness, spatial arrangement
 Extending the classification to 3 classes of Absent,
Typical, and Atypical
 Color of the of surrounding network in blue channel
 Thickness and irregularity of the network
 Modifying the method to find other dermoscopic
structures and patterns
25
Questions?
Thank you
26
Conclusion
 A novel graph-based method for classifying and
visualizing pigment networks.
 Evaluating its ability to classify and visualize real
dermoscopic images
 The accuracy of the method is 92.6% (classifying
images to Absent and Present)
27
Previous Work
 Comparing our results to previous methods:
 Anantha et al. “Detection of pigment network in
dermatoscopy images using texture analysis” , 2004,
Accuracy: 80%
 Betta et al. “Dermoscopic image-analysis system: estimation
of atypical pigment network and atypical vascular pattern”,
2006, Recall:50% , Precision: 100%, F-measure: 66.66%
 Our method: Accuracy: 92.6%
28
Original
Laplacian of Gaussian
Image to Graph
Pigment Network
Cyclic Subgraphs
Present
Classification
29