Mammography is currently the dominant imaging modality for the early detection of breast cancer. However, its robustness in distinguishing malignancy is relatively low, resulting in a large number of unnecessary biopsies. A computer-aided diagnosis (CAD) scheme, capable of visually justifying its results, is expected to aid the decision made by radiologists. Content-based image retrieval (CBIR) accounts for a promising paradigm in this direction. Facing this challenge, we introduce a CBIR scheme that utilizes the extracted features as input to a support vector machine (SVM) ensemble. The final features used for CBIR comprise the participation value of each SVM. The retrieval performance of the proposed scheme has been evaluated quantitatively on the basis of the standard measures. In the experiments, a set of 90 mammograms is used, derived from a widely adopted digital database for screening mammography. The experimental results show the improved performance of the proposed scheme.
The use of content-based image retrieval (CBIR) schemes for computer-aided
diagnosis (CAD) has been intensively investigated in the last decade [
There is only a limited amount of works devoted to CBIR-based CAD for
breast masses in mammograms, although mammographic CAD is one of a mature
and widely adopted CAD type [
All the above works evaluated retrieval performance on the basis of discriminant capability between benign and malignant cases. It can be argued that a CBIR scheme is expected to retrieve cases on the basis of visual similarity, since by its own nature it cannot take into account accompanying clinical data. In a real clinical setting, the results of CBIR could be jointly assessed with all such data in the context of an integrated CAD scheme. Finally, it can be observed that all CBIR methods presented were based on simple similarity measures, which cannot optimally exploit the distribution of mammogram ROIs in the feature space.
Dataset Query Image
Feature Calculation
OFFLINE
CalculatetheParticipation ValuefromeachTrained
SVM
In this work, we present a novel CBIR scheme, which utilizes a support vector
machine (SVM) ensemble. The corresponding SVMs are capable of optimally
exploiting the distribution of input samples in the feature space on the basis of
breast imaging-reporting and data system (BI-RADS) classifications of breast
masses [
The remainder of the paper is organized as follows: Section 2 describes the architecture of the proposed system. The experimental evaluation of the proposed CBIR scheme is presented in Section 3. Finally, conclusions and future perspectives of this work are discussed in Section 4. 2
pipeline. Since those two stages are out of the scope of the proposed system, there will be no further discussion about the methodology used. For our experiments, the segmentation information is taken from the ground truth corpora.
Initially, the following features are extracted from the mass boundaries:
1. Solidity: Fsolidity = A/H, where A and H denote the areas of the shape and
its corresponding convex hull, respectively.
2. Compactness: Fcompactness = 1 − (4πA/P 2), where A and P denote the area
and the perimeter of the shape, respectively.
3. Discrete Fourier Transformation (DFT) coefficients of the Normalized
Radial Length (NRL) function. The Radial Length Function corresponds to
the distance of each contour point (t) from the mass centroid (xc, yc): r(t) =
p(x(t) − xc)2 + (y(t) − yc)2. This function is sampled to a fixed number of
points (256 in this work) and is normalized before its DFT computation.
Thereafter, the above feature set is supplied to three different trained SVMs
that correspond to three classes of breast masses, based on BI-RADS [
The Support Vector Machines (SVMs) [
Instead of utilizing the sign of the SVM decision function, we propose to
normalize it, based on [
max n 1 − max
o
1n+ 311ef(x) , 1+ 13 e1−f(x) if f (x) > 0
1+ 131ef(x) , 1+ 13 e1−f(x) o if f (x) < 0
(1)
where f (x) denotes the SVM decision function. The output of the Eq. 1
represents the membership value of the data x to the corresponding class and ranges
in the interval [
In this study, 90 regions of interest (ROI) were used, extracted from various
mammograms of DDSM, which contain masses. Each case is accompanied with
ground-truth delineations and additional information, such as the biopsy-proven
pathology of the lesion, its shape and margin types, the overall breast density
and the assessment, of an expert radiologist [
Two performance evaluation metrics are employed, which measure the system’s ability to retrieve masses of similar margin type to the query. The first one is the Precision at Top 5 Retrieved items (P@5), which defines how successfully the algorithms produce relevant results to the first 5 position of the ranking list. The second metric used, is the Mean Average Precision (MAP) which is a typical measure for the performance of information retrieval systems and it is defined as the average of the precision value obtained after each relevant retrieved item.
The BI-RADS SVMs used a radial basis function (RBF) kernel, they were trained by the 2/3 and evaluated with the remainder portion of the dataset.
The proposed method was evaluated against a typical, unsupervised, stateof-the-art retrieval system employing the euclidean distance, calculated directly from the features instead of the participation values from each SVM. Comparative results are presented in Table 1 and show that the proposed method outperformed the typical unsupervised euclidean-based retrieval system. evaluated on the basis of BI-RADS classifications of breast masses. The used dataset is formed by 90 cases of the DDSM, which is a dataset widely adopted by the medical community. The experimental results lead to the conclusion that the proposed CBIR scheme outperforms standard euclidean-based retrieval, while greatly reducing the feature vector dimension and, consequently, the computational cost.
Future perspectives of this work include: 1) the integration of the proposed CBIR scheme within the context of a mammographic CAD system, which will also consider accompanying clinical and textual data, 2) the development of a similar CBIR scheme to facilitate CAD of breast microcalcifications.
This work is funded by the Hellenic Republic, Ministry of Education and Religious Affairs, General Secretariat of Research and Technology (GSRT), and particularly the National Programme “SYNERGASIA 2011” (11SYN 10 1546) in the National Strategic Reference Framework (NSRF) 2007-2013.