Sparse feature analysis for detection of clustered microcalcifications in mammogram images
1. SPARSE FEATURE ANALYSIS FOR DETECTION OF CLUSTERED
MICROCALCIFICATIONS IN MAMMOGRAM IMAGES
Wonyong Eom, Wesley De Neve, and Yong Man Ro
Image and Video Systems Lab
Korea Advanced Institute of Science and Technology (KAIST)
Daejeon, South Korea
e-mail: ymro@ee.kaist.ac.kr website: http://ivylab.kaist.ac.kr
- Features
I. INTRODUCTION Table. 1. Feature types and dimensionality
- Observation Features Dimension
• Computer-aided detection (CADe) of clustered microcalcifications First Order Statistics (FOS) 16
(MCs) in mammogram image is one of the most effective tools for Rotation Invariant Moment (RIM) 16
detecting early-stage breast cancer. Spatial Gray Level Difference (SGLD) 52
Gray Level Run Length (GLRL) 20
• A number of CADe systems have recently introduced the use of
Laws’ Texture Features (LAW) 250
sparse representation based classification (SRC). Uniform Local Binary Patterns (LBP) 118
- Problem statement
- Evaluation
- Few attempts have thus far been made to achieve an in-depth
• the area under the ROC curve (AUC)
understanding of the influence of SRC on the effectiveness of these
• the sparsity concentration of the true class (SCTC)
CADe systems.
- Contributions δ T ( x) 1
SCTC (x) = ∈ [0, 1],
- We compare and analyze the influence of commonly used features, x1
different feature combinations and different dictionary construction
where x represents the sparse coefficient vector, and where δT
techniques on the effectiveness of an SRC-based CADe systems.
denotes the true class part of x
2. Feature comparison
II. METHOD Table. 2. Effectiveness of SRC-based detection of MC according the feature used
1. Dictionary construction Feature AUC SCTC
FOS 0.8885 0.7923
f1 RIM 0.8546 0.7855
SGLD 0.8986 0.8011
GLRL 0.8814 0.7891
…
f
[[
Feature LAW 0.9403 0.8402
f 2 normalization/
…
Feature
…
ROI LBP 0.9322 0.8168
detection extraction concatenation
…
… 3. Feature combination comparison
…
…
… … Table. 3. Effectiveness of SRC-based detection of MCs according to the feature
…
ROI combination used
fi
Feature Combination AUC SCTC
…
…
FOS+RIM 0.9047 0.8112
…
Mammogram
Dictionary SGLD+GLRL+LAW+LBP 0.9374 0.8324
Fig. 1. Creating a dictionary of image features FOS+LAW 0.9326 0.8315
2. Sparse representation based classification LAW+LBP 0.9483 0.8392
All features 0.9525 0.8401
Malignant Normal 4. Effectiveness of dictionary size
ROIs ROIs - We gradually reduced the number of dictionary atoms from 90% of the
available samples to 50% of the available samples.
- This implied that smaller dictionaries were subsets of larger dictionaries.
Feature extraction Test sample Dictionary Sparse
coefficient
- We made use of Laws’ texture feature.
vector
Residual for Residual for
malignant ROI normal ROI
Dictionary
Classify as
Fig. 2. Sparse representation based classification method
III. EXPERIMENTS
1. Experimental setup
- Dataset
• We made use of 180 malignancy-containing X-ray images randomly
taken from the Digital Database for Screening Mammography(DDSM).
• From these images, we obtain 434 malignant ROIs and 2556 normal Fig. 3. Different ROC curves according to the number of dictionary atoms
tissue region using a contrast-based method for candidate region IV. CONCLUSIONS
detection.
- Our experimental results show that the use of texture features is more
• We used 10 percent of the positive samples and 10 percent of the
effective than the use of shape and morphology features.
negative samples for the purpose of testing, while the remaining
- SRC based MCs detection with LAW and the combination LAW+LBP is
samples were used for the purpose of dictionary construction. We
highly promising.
repeated this ten times with different test sets, and then averaged
- Our experimental results show that the more atoms in the dictionary,
the results obtained for each run in order to compute the final
the higher the discriminative power of SRC-based CADe system.
results.
International Forum on Medical Imaging in Asia (IFMIA), November 2012, Daejeon(Korea)