26
A two-stage SVM-based mammographic CBIR for CADx L. TSOCHATZIDIS 1 A. KARAHALIOU 2 K. ZAGORIS 1 S. SKIADOPOULOS 2 N. ARIKIDIS 2 L. COSTARIDOU 2 I. PRATIKAKIS 1 University of Patras 2 School of Medicine Department of Medical Physics Democritus University of Thrace 1 Department of Electrical and Computer Engineering Visual Computing Group TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 03/03/2022 1

A two stage svm-based mammographic cbir for ca dx

Embed Size (px)

Citation preview

Page 1: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 1

A two-stage SVM-based mammographic CBIR for CADxL. TSOCHATZIDIS 1

A. KARAHALIOU 2

K. ZAGORIS 1

S. SK IADOPOULOS 2

N. ARIK IDIS 2

L. COSTARIDOU 2

I . PRATIKAKIS 1

University of Patras 2School of MedicineDepartment of Medical Physics

Democritus University of Thrace 1Department of Electrical and Computer EngineeringVisual Computing Group

Page 2: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 2

CADx in MammographyMammography is a dominant imaging modality for early detection of breast cancer

Often, diagnosis leads to unnecessary biopsies

Two types of CADx:• Single-stage: Classification schemes for benign-

malignant discrimination• Two-stage: Content-Based Image Retrieval that

feeds the diagnosis step which discriminates between benign and malignant

Page 3: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 3

Proposed CBIR-CAD SystemCAD system that incorporates a CBIR step and a decision step

Retrieves similar images based on low-level image features

Margin specific CBIR

Diagnosis is based on the ranked lists produced by CBIR

Provides visual aid and enables consultation of previous cases, leading to increased confidence into incorporating CAD-cued results

Page 4: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 4

Margin-type classesCircumscribed

Spiculated

Microlobulated

Ill defined (+Obscured)

Page 5: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 5

CBIR-CAD’s pipeline

BENIGN / MALIGNANT

Page 6: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 6

Semi-automatic SegmentationThe Dijkstra’s shortest path algorithm is exploited to obtain the optimal path between sequential pairs of landmark points upon mass boundaries

A new cost function is proposed to avoid background correction techniques that may deform mass contour and introduce additional adjustment parameters

Arikidis, N., Skiadopoulos, S., Karahaliou, A., Kazantzi, A., Vassiou, K., Tsochatzidis, L., Pratikakis, I., Costaridou, L.: Shortest paths of mass contour estimates in mammography. In: MICCAI-BIA 2015

Page 7: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 7

CBIR Architecture

Page 8: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 8

Feature Extraction – Global ShapeSolidity factor: The degree that the shape deviates from its convex hull

Compactness factor: The degree that a shape deviates from a perfect circle

Page 9: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 9

Feature Extraction – Global Shape

Circumscribed Microlobulated Spiculated Ill-defined

Compactness=Solidity=0.99

Compactness=Solidity=0.92

Compactness=Solidity=0.32

Compactness=Solidity=0.93

Page 10: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 10

Feature Extraction – DFT of NRLNormalized 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

Page 11: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 11

Feature Extraction - Texture-basedRubber Band Straightening Transform (RBST)

•Unfolding the ribbon around the contour as a flat image

•RBST Column line segment normal to the contour

•RBST Row iso-distant to the contour paths

•Intensity profiles at every contour point along a line segment normal to the contour

Page 12: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 12

Feature Extraction - Texture-based

RBST Image

Sobel gradient magnitude operator

Detected edge points

Page 13: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 13

Feature Extraction - Texture-based

Page 14: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 14

Feature Extraction - Texture-based

Extracted Features

•Distance between edge points of consecutive columns

•Distance between edge points and middle row of RBST image

•Magnitude of gradient on y-axis

•Gradient orientation divergence from vertical direction

•Acutance (The sum of the difference of gray-level values between pixels that are iso-distant from either sides of the contour)

Mean and SD value of the above functions

Page 15: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 15

Feature Extraction - Texture-based

Feature Name Circumscribed Ill-defined

Avg. dist. edge points 0.036 0.747

Avg. dist center row 0.323 0.865

SD dist. edge points 0.105 0.879

SD dist center row 0.077 0.952

Page 16: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 16

CBIR Architecture

Page 17: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 17

The SVM Layer – Support Vector MachinesBinary 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:

Page 18: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 18

The SVM Layer – Structure

An ensemble of binary SVM classifiers is employed

One SVM for each class – Four SVMs in total

Each SVM outputs the participation level of a sample in the corresponding class

Page 19: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 19

CBIR Architecture

Page 20: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 20

The Diagnosis StageGOAL: Provide the likelihood of malignancy for a query case.

•Based on the K most similar ROIs retrieved

•Similarity between query and an item: ,

•Two decisions indices investigated:

Page 21: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 21

Experimental ResultsExperiments on a dataset of total 400 mammograms from DDSM

Precise contour delineation from expert radiologist

CC and MLO views are treated independently

5-fold cross validation

Grid search for SVM and kernel parameters

Page 22: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 22

Experimental Results – Evaluation metricsPrecision at N (P@R): The percentage of correct images at the top-R places of the rank list

Mean Average Precision (MAP): Measures the overall performance of a query

Page 23: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 23

Experimental Results - CBIRClasses P@R MAP

Circumscribed

Microlobulated

Spiculated

Ill-defined

Average

Page 24: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 24

Area under ROC curve (AUC)The Receiver Operating Characteristic (ROC) curve illustrates the performance of a binary classifier as its discrimination threshold is varied

The curve is created by plotting the true positive rate (sensitivity) against the false positive rate (1-specificity) at various threshold settings

The AUC of a classifier is equivalent to the probability that the classifier will rank a randomly chosen positive instance higher than a randomly chosen negative instance

Page 25: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 25

Experimental Results - Decision

3 4 5 6 7 8 9 10 11 12 13 14 150.74

0.75

0.76

0.77

0.78

0.79

0.8

0.81

0.82

Classification performance of D1 and D2 in terms of Az index.

D1 D2

Maximum using D2 for K=13 ranked items

Page 26: A two stage svm-based mammographic cbir for ca dx

03/05/2023 TSOCHATZIDIS ET AL. A TWO-STAGE SVM-BASED MAMMOGRAPHIC CBIR FOR CADX, MICCAI-BIA 2015 26

ConclusionsTwo-stage CBIR-CAD:• Margin-specific CBIR stage• Diagnosis stage

Incorporation of training into the feature extraction (SVM ensemble)

High-performance for spiculated and microlobulated masses

Lack of standard datasets leads to difficulty in comparison between methods

Future efforts:• Performance improvement for ill-defined masses• Feature selection for each SVM independently• Introduction of weights in decision calculation modifying the significance of each retrieved ROIs

• Use of relevance feedback mechanism to improve performance