MammoSys: A content-based image retrieval system using breast density patterns

TLDR: A content-based image retrieval system designed to retrieve mammographies from large medical image database based on breast density, and integrated to the database of the Image Retrieval in Medical Applications (IRMA) project, that provides images with classification ground truth.

About: This article is published in Computer Methods and Programs in Biomedicine.The article was published on 2010-09-01 and is currently open access. It has received 88 citations till now. The article focuses on the topics: Content-based image retrieval & Automatic image annotation.

Figures

Table 2 – Execution time, in seconds, for the CBIR

Table 1 – Summary of published works based on the analysis of breast density, with the developed system, number of

Fig. 3 – Precision × recall curve for the average precision using the first five principal components for breast density

Table 3 – Average precision for the selected d first

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Mammosys: a content-based image retrieval system using breast density patterns" ?

In this paper, the authors present a content-based image retrieval system designed to retrieve mammographies from large medical image database. The system is developed based on breast density, according to the four categories defined by the American College of Radiology, and is integrated to the database of the Image Retrieval in Medical Applications ( IRMA ) project, that provides images with classification ground truth. 

Q2. What are the future works in "Mammosys: a content-based image retrieval system using breast density patterns" ?

Future works may additionally consider other patterns for retrieval, such as breast lesions, masses and calcifications, characterized by size and shape. Breast lesions may be used together with breast density, providing a more instructive CBIR system for radiologists, as more information may become available to support diagnosis. 

Q3. What is the optimal projection axis Xopt?

The optimal projection axis Xopt is the unitary vector that maximizes J(X), i.e. the eigenvector of G corresponding to the largest eigenvalue. 

Q4. What is the method for calculating the distance between two features of an image?

The principal component vectors obtained are used to form an m × d matrix L = [YT1 , Y T 2 , . . . , Y T k ], which is called the feature matrix or feature image of the image A. Some works employed 2DPCA technique for face and palmprint representation. 

Q5. What are the main challenges in the development of CBIR systems?

The appropriate characterization of images together with the storage and management of the large amount of images produced by hospitals and medical centers are a main challenge in the development of CBIR systems. 

Q6. What is the significance of the histograms?

Histograms were used for the characterization of breast ensity in a set of 195 mammographies at the Medical Cener of Pittsburgh by Wang et al. [13], in order to automatically valuate breast density according to BI-RADS categories. 

Q7. What processor was used for image retrieval?

Feature extraction was performed on an Intel Core2Quad 2.66 GHz processor with 8 GB of RAM under Microsoft Windows operating system and image retrieval was executed on an Intel Core2Duo 2 GHz processor with 3 GB of RAM, also under Microsoft Windows operating system. 

Q8. What is the method for characterization of breast density?

In the proposed system, ROIs containing only breast density were characterized using 2DPCA, a novel and promising method for the characterization texture in lowdimensional feature spaces. 

Q9. How many subimages were used to test the efficiency of the proposed method?

using the PolyU palmprint database (2004), they used 600 subimages of size 128 × 128 pixels to test the efficiency of the proposed method. 

Q10. What is the way to extract the image from a database?

The query image goes through the process of feature extraction in order to be compared to the feature vectors of all images stored in the database. 

Q11. What is the classifier for image retrieval?

2DPCA(w/o3)PCA consumed less time for the extraction of the features and also obtained the highest accuracy rate – 99.27% – using a classifier proposed by the authors, a modified modular neural network (MNN) classifier. 

Q12. What is the first approach to the problem of multiple lasses?

The first approach reduces the problem of multiple lasses to a set of binary problems, using methods of decomosition one by class (one against all) and the separation of lasses two by two (one against one). 

Q13. What are the optimal projections vectors of 2DPCA?

These optimal projections vectors of 2DPCA, X1, . . . , Xd are used for feature extraction, where d corresponds to the number of selected eigenvalues. 

Q14. What is the difference between two-dimensional principal component analysis and PCA?

Two-dimensional principal component analysis overcomes principal component analysis (PCA) as it is simpler and more straightforward to use for image feature extraction since 2DPCA is directly applied to the image matrix. 

Q15. What is the simplest way to retrieve breast density?

In this paper the authors presented a CBIR system that uses breast density as a pattern for image retrieval and is able to aid radiologists in their diagnosis. 

Q16. What is the complex task in the rocess?

The defiition of a set of features, capable to effectively describe each egion of the image, is one of the most complex tasks in the rocess. 

Q17. What is the covariance matrix of the feature vectors of the training examples?

Sx denotes the covariance matrix of the projected feature vectors of the training examples and tr(Sx) denotes the trace of Sx:tr(Sx) = XT[E(A − EA)T(A − EA)]X (3)The image covariance matrix G of an image A can be defined as:G = E[(A − EA)T(A − EA)] 

Q18. What is the owest value of the rst five principal components?

The owest values are close to zero and can be considered insuffiient to properly characterize the images, explaining the fact hat average precisions decrease as the number of principal omponents increases.