�ݺ�ߣ

SVM-based CBIR of Breast
Masses on Mammograms
L. TSOCHATZIDIS
K. ZAGORIS
M. SAVELONAS
I. PRATIKAKIS
Democritus University of Thrace
Department of Electrical and Computer Engineering
Visual Computing Group
August 17, 2014
European Conference on Artificial Intelligence (ECAI) 2014 ,
Workshop on Artificial Intelligence and Assistive Medicine (AI-AM)
1

Mammography
 Diagnostic and screening tool of
breasts
 Dominant imaging modality for early
detection of breast cancer
 Breast cancer appears as a mass
and/or microcalcifications
 The diagnosis is difficult that leads to
unnecessary biopsies
August 17, 2014
Workshop on Artificial Intelligence and Assistive Medicine
(AI-AM)
2

Computer Aided Decision (CAD)
 Consists of two sub-categories:
 Systems for detecting an abnormality - Computer Aided Detection (CADe)
 Systems for diagnosing an abnormality - Computer Aided Diagnosis (CADx)
 CAD systems usually employ classification schemes for benign-malignant
discrimination
August 17, 2014
(AI-AM)
3

CBIR-CAD Systems
 CAD systems that incorporate a CBIR step prior to decision
 Retrieve similar images based on low-level image features
 Provide visual aid, enables consulting previous cases, leading to
increased confidence into incorporating CAD-cued results
August 17, 2014
(AI-AM)
4

Problem Definition – Classes (1)
 Circumscribed  Smooth and highly convex boundary
 Well-defined margin
 Low probability of malignancy
August 17, 2014
(AI-AM)
5

 Microlobulated  Rough and bumpy boundary
 The overall shape is retained
 Medium to high probability of
malignancy
August 17, 2014
(AI-AM)
6

 Spiculated  Margin with large protrusions and not
clearly defined
 The overall shape becomes irregular
 Highly suggestive of malignancy
August 17, 2014
(AI-AM)
7

CBIR-CAD’s pipeline
August 17, 2014
(AI-AM)
Lesion
Segmentation
Mass
Detection
CBIR Diagnosis
Input
Image
BENIGN
8

CBIR Architecture
August 17, 2014
(AI-AM)
9

Feature Extraction – Global Shape (1)
 Solidity factor: The degree that the shape deviates from its convex hull
𝑆𝑜𝑙𝑖𝑑𝑖𝑡𝑦 =
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑚𝑎𝑠𝑠 (𝐴)
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑖𝑡𝑠 𝑐𝑜𝑛𝑣𝑒𝑥 ℎ𝑢𝑙𝑙 (𝐻)
 Compactness factor: The degree that a shape deviates from a perfect
circle
𝐶𝑜𝑚𝑝𝑎𝑐𝑡𝑛𝑒𝑠𝑠 = 1 −
4𝜋𝐴2
𝑃2
August 17, 2014
(AI-AM)
10

Feature Calculation – Global Shape
(2)
August 17, 2014
(AI-AM)
Circumscribed Microlobulated Spiculated
11

Feature Extraction – DFT of NRL
 Normalized Radial Length Function
1. The distance of each contour point
to the shape’s center of gravity
2. Normalized by the average radial
length
3. Computation of Discrete Fourier
Transform coefficients
August 17, 2014
(AI-AM)
12

CBIR Architecture
August 17, 2014
(AI-AM)
13

The SVM Layer – Support Vector
Machines (1)
 Binary Linear Classifiers
 For non-linear problems: Projection of
samples to a higher dimensionality
space.
 Finds a hyper-plane that optimally
separates the two classes
 Decision function:
𝑓 𝑥 = 𝑠𝑖𝑔𝑛(𝑤 ∙ 𝑥 + 𝑏)
August 17, 2014
(AI-AM)
14

The SVM Layer – Structure
 An ensemble of binary SVM classifiers
is employed
 One SVM for each class – Three SVMs
in total
 Each SVM outputs the participation
level of a sample in the corresponding
class
August 17, 2014
(AI-AM)
15

The SVM Layer – Participation value
Computation
 SVM Decision Function
𝑓 𝑥 = 𝒘 ∙ 𝒙 + 𝑏
 Proposed normalization
𝑅 𝑥 =
𝑚𝑎𝑥
1
1 +
1
3
𝑒 𝑓(𝑥)
,
1
1 +
1
3
𝑒−𝑓(𝑥)
, 𝑖𝑓 𝑓 𝑥 > 0
1 − 𝑚𝑎𝑥
1
1 +
1
3
𝑒 𝑓 𝑥
,
1
1 +
1
3
𝑒−𝑓 𝑥
, 𝑖𝑓 𝑓 𝑥 < 0
August 17, 2014
(AI-AM)
16

Experimental Results
 Experiments on a dataset of total 90 mammograms (CC views) from DDSM
 Manual contour delination from expert radiologist
 Equal number of mammograms from each class
 The 2/3 of dataset was used for the SVM training
 The Rest 1/3 was used as test set
 Comparison between proposed method and the typical, unsupervised
one.
August 17, 2014
(AI-AM)
17

Experimental Results – Evaluation
metrics
 Precision at N (P@N): The percentage of correct images at the top-N
places of the rank list (N=5)
 Mean Average Precision (MAP): Measures the overall performance of a
query
𝐴𝑃 =
𝑘=1
𝑛
𝑃@𝑘 ∗ 𝑟𝑒𝑙 𝑘
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠
𝑟𝑒𝑙 𝑘 =
1, 𝑖𝑓 𝑘_th 𝑖𝑚𝑎𝑔𝑒 𝑖𝑠 𝑐𝑜𝑟𝑟𝑒𝑐𝑡
0, 𝑒𝑙𝑠𝑒
𝑀𝐴𝑃 =
1
𝑁
𝑁
𝐴𝑃
August 17, 2014
(AI-AM)
18

Experimental Results
Classes Unsupervised CBIR Supervised CBIR
P@5 MAP P@5 MAP
Circumscribed 0.90 0.91 0.90 0.92
Microlobulated 0.71 0.72 0.71 0.76
Spiculated 0.65 0.61 0.74 0.73
Average 0.75 0.74 0.78 0.80
August 17, 2014
(AI-AM)
19

Experimental Results – Circumscribed
August 17, 2014
(AI-AM)
20

Experimental Results – Microlobulated
August 17, 2014
(AI-AM)
21

Experimental Results – Spiculated
August 17, 2014
(AI-AM)
22

Conclusions
 CBIR system for retrieval of masses on mammograms
 Evaluation showed that the features have high discriminant ability
 The supervised CBIR offers enhanced results as compared to the
unsupervised one.
 The final vectors used are very small compared to the initial feature
vectors
August 17, 2014
(AI-AM)
23

Future Work
 Test the CBIR scheme for microcalcifications
 Integration of CBIR within the context of a complete mammographic CAD
system
August 17, 2014
(AI-AM)
24

Thank you!
Any questions?
August 17, 2014
(AI-AM)
25

�ݺ�ߣ

Svm based cbir of breast masses on mammograms

Recommended

More Related Content

Similar to Svm based cbir of breast masses on mammograms (20)

Recently uploaded (20)

Svm based cbir of breast masses on mammograms