A new multi-expert decision combination algorithm and its application to the detection of circumscribed masses in digital mammograms

doi:10.1016/S0031-3203(00)00088-1

Pattern Recognition

Volume 34, Issue 8, August 2001, Pages 1527-1537

https://doi.org/10.1016/S0031-3203(00)00088-1 Get rights and content

Abstract

A new multiple expert fusion algorithm is introduced, designated the “augmented behaviour-knowledge space method”. Most existing multiple expert classification methods rely on a large training dataset in order to be properly utilised. The proposed method effectively overcomes this problem as it exploits the confidence levels of the decisions of each classifier. It will be shown that this new approach is advantageous when small datasets are available, and this is illustrated in its application to the detection of circumscribed masses in digital mammograms, with very encouraging results.

Introduction

Practical applications in pattern recognition have been shown to benefit from the combination of the decisions of different algorithms and techniques (“experts”) since classifiers with different internal structures can complement each other [1], [2], [3]. Such multiple expert strategies are likely to deliver more robust decisions than individual classifiers working alone. As a result, it is now common to adopt a variety of decision fusion algorithms to combine the individual expert decisions. The decision combination process has to merge the individual decisions in such a way that the final classification improves the classification profile of any of the individual experts.

This paper proposes a new method of classifier fusion, the augmented behaviour knowledge space method (ABKS), which is an extended and enhanced implementation of the established behaviour knowledge space method (BKS) technique [4], but which specifically addresses a principal weakness in that technique which renders it unsuitable for many important practical applications. Specifically, the proposed ABKS technique is applicable to situations where large training sets are not available, thus avoiding the degree of unreliability which often leads to a deterioration in performance when the classifier is not correctly initialised.

The proposed method is applied to the difficult task of the detection of circumscribed masses in digital mammograms and is shown to achieve very encouraging results. The proposed technique is particularly well suited to the problem of the detection of circumscribed masses, where only small image databases are often available. Due to the sensitive and critical nature of this task domain, it is vital to be able to achieve a robust and reliable decision under these conditions.

Section snippets

Overview of operation

Many practical techniques for classifier fusion are based on a majority voting approach. Most of these methods treat each classifier equally, or derive information useful for the combination stage from the confusion matrices of each classifier. In order to do this, the assumption has to be made that the decisions of all the classifiers are independent [5]. The ABKS method specifically avoids this (often inaccurate) assumption as it derives its weights from the knowledge space, which

Mammography application

The ABKS technique described in the previous sections has been adopted to investigate a complete screening system to detect cancerous lesions in mammograms. Breast cancer is a leading cause of fatality in women, with approximately 1 in 12 women affected by the disease during their lifetime [7]. Mass screening of women using X-ray mammography is currently the most effective method of early detection of the disease, which is essential for successful treatment. Currently, mammograms are visually

Discussion

Based on the results generated with the ABKS technique, some general conclusions can be drawn. In the ABKS evaluation method described in Section 2.4, a 5×5 ABKS space is used to select two classes. This means that there are 2⁵ (32) focal units. The ABKS space was trained using 20 ABNORMAL and 20 NORMAL samples extracted from the defined training data set. The results reported in Section 2.4 were obtained using the remaining samples of the data set, over a set of 10 runs. At each run there was

Summary

This paper has introduced and evaluated a new decision fusion approach designated the Augmented Behaviour Knowledge Space (ABKS) method. This method has been shown to be beneficial in pattern recognition problems when there is a relatively small training dataset available. The proposed method thus provides a reliable method of addressing problems encountered in other classifier fusion schemes and provides a more reliable method of classifier integration in many applications. It also needs to be

About the Author—ALEX CONSTANTINIDIS graduated with a B.Sc. Honours degree in Digital Electronics from the University of Kent in UK in 1994. Following graduation he embarked on a research programme studying for a Ph.D. degree, also at the University of Kent, which he will complete during 2000. His research interests are wide but include principally image recognition, medical image processing and computer security applications.

References (18)

L Wang et al.
Texture classification using texture spectrum
Pattern Recognition
(1990)
M.C Fairhurst et al.
A generalised approach to the recognition of structurally similar handwritten characters
IEE Proc. Vision Image Signal Process.
(1997)
A.F.R Rahman et al.
Enhancing multiple expert decision combination strategies through exploitation of a priori information sources
IEE Proc. Vision Image and Signal Process.
(1999)
J. Kittler, A. Hojjatoleslami, T. Windeatt, Weighting factors in multiple expert fusion, British Machine Vision...
Y.S Huang et al.
A method of combining multiple experts for the recognition of unconstrained handwritten numerals
In IEEE Trans. Pattern Anal. Mach. Intell.
(1995)
L Lam et al.
Application of majority voting to pattern recognitionan analysis of its behavior and performance
IEEE Trans. Pattern Anal. Mach. Intell.
(1997)
Y.S Huang et al.
A method of combining multiple experts for the recognition of unconstrained handwritten numerals
In IEEE Trans. Pattern Anal. Mach. Intell.
(1995)
D Brzakovic et al.
An approach to automated detection of tumours in mammography
IEEE Trans. Med. Imaging
(1990)
A.S. Constantinidis, M.C. Fairhurst, F. Deravi, M. Hanson, C.P. Wells, C. Chapman-Jones, Evaluating classification...

There are more references available in the full text version of this article.

Cited by (37)

Applying a new quantitative image analysis scheme based on global mammographic features to assist diagnosis of breast cancer
2019, Computer Methods and Programs in Biomedicine
Citation Excerpt :
Last, similar to our previous study, which demonstrated that fusion of the complementary information of global (case-based) and regional (lesion-based) mammographic image features had potential to significantly improve CAD performance in detecting suspicious lesions without increase of false-positive rates [34], we will investigate the optimal fusion method to fuse the classification results of this global feature based scheme with the lesion-based scheme [35] to more accurately and robustly classify between malignant and benign mammographic cases in the future.
This study aims to develop and evaluate a unique global mammographic image feature analysis scheme to predict likelihood of a case depicting the detected suspicious breast mass being malignant for breast cancer.
From the entire breast area depicting on the mammograms, 59 features were initially computed to characterize the breast tissue properties at both spatial and frequency domain. Given that each case consists of two cranio-caudal and two medio-lateral oblique view images of left and right breasts, two feature pools were built, which contain the computed features from either two positive images of one breast or all the four images of two breasts. Next, for each feature pool, a particle swarm optimization (PSO) method was applied to determine the optimal feature cluster followed by training a support vector machine (SVM) classifier to generate a final score for predicting likelihood of the case being malignant. To test the scheme, we assembled a dataset involving 275 patients who had biopsy due to the suspicious findings on mammograms. Among them, 134 are malignant and 141 are benign. A ten-fold cross validation method was used to train and test the scheme.
The classification performance levels measured by the areas under ROC curves are 0.79 ± 0.07 and 0.75 ± 0.08 when applying the SVM classifiers trained using image features computed from two-view and four-view images, respectively.
This study demonstrates feasibility of developing a new global mammographic image feature analysis-based scheme to predict the likelihood of case being malignant without lesion segmentation.
An automated approach to the classification of impact spatter and cast-off bloodstain patterns
2018, Forensic Science International
Citation Excerpt :
The design of a pattern recognition system generally consists of the following stages: defining pattern classes, data collection, selecting the distinguishing features (referred to as feature selection), specifying classification algorithms and estimating the classification error [14]. Over the years, pattern recognition systems have evolved to become valuable tools that are used to organise or retrieve vast amounts of electronic data [15,16], develop computer-aided diagnosis systems for the detection of disease [17–19] or to identify signature characteristics of fluids [20]. It has become evident, that forensic disciplines are moving away from manual methods of analysis towards more automated computer-based methods of pattern recognition.
In the forensic discipline of bloodstain pattern analysis, it has been suggested that there is a blurred boundary between characterising the features of a bloodstain pattern and determining the mechanism(s) that led to its deposition. This study proposes that bloodstain pattern classification can become a distinct and logical process by implementing an automated approach. To do this, an automated bloodstain pattern recognition system was developed to enable the distinction of two types of spatter bloodstain patterns. First, global pattern features based on common bloodstain pattern properties were extracted from laboratory-generated impact spatter and cast-off bloodstain patterns. Following this, automated feature selection methods were used to identify the combination of features that best distinguished the two bloodstain pattern types. This eventually led to the training and testing of a Fisher quadratic discriminant classifier using separate subsets of the generated bloodstain patterns. When applied to the training dataset, a 100% classification precision resulted. An independent dataset comprising of bloodstain patterns generated on paint and wallpaper substrates were used to validate the performance of the classifier. An error rate of 2% was obtained when the classifier was applied to these bloodstain patterns. This automated bloodstain pattern recognition system offers considerable promise as an objective classification methodology which up to now, the discipline has lacked. With further refinement, including testing it over a wider range of bloodstain patterns, it could provide valuable quantitative data to support analysts in their task of classifying bloodstain patterns.
Classifier ensemble generation and selection with multiple feature representations for classification applications in computer-aided detection and diagnosis on mammography
2016, Expert Systems with Applications
This paper presents a novel ensemble classifier framework for improved classification of mammographic lesions in Computer-aided Detection (CADe) and Diagnosis (CADx) systems. Compared to previously developed classification techniques in mammography, the main novelty of proposed method is twofold: (1) the “combined use” of different feature representations (of the same instance) and data resampling to generate more diverse and accurate base classifiers as ensemble members and (2) the incorporation of a novel “ensemble selection” mechanism to further maximize the overall classification performance. In addition, as opposed to conventional ensemble learning, our proposed ensemble framework has the advantage of working well with both weak and strong classifiers, extensively used in mammography CADe and/or CADx systems. Extensive experiments have been performed using benchmark mammogram dataset to test the proposed method on two classification applications: (1) false-positive (FP) reduction using classification between masses and normal tissues, and (2) diagnosis using classification between malignant and benign masses. Results showed promising results that the proposed method (area under the ROC curve (AUC) of 0.932 and 0.878, each obtained for the aforementioned two classification applications, respectively) impressively outperforms (by an order of magnitude) the most commonly used single neural network (AUC = 0.819 and AUC =0.754) and support vector machine (AUC = 0.849 and AUC = 0.773) based classification approaches. In addition, the feasibility of our method has been successfully demonstrated by comparing other state-of-the-art ensemble classification techniques such as Gentle AdaBoost and Random Forest learning algorithms.
A decision-making algorithm for automatic flow pattern identification in high-speed imaging
2014, Expert Systems with Applications
A digital image processing algorithm was developed to identify flow patterns in high speed imaging. This numerical tool allows to quantify the fluid dynamic features in compressible flows of relevance in aerospace and space related applications. This technique was demonstrated in a harsh environment with poor image quality and illumination fluctuations. This original pattern recognition tool is based on image binarization and object identification. The geometrical properties of the detected elements are obtained by measuring the characteristics of each object in the binary image. In case of multiple shock waves or shock bifurcations, a “decision-making” algorithm chooses the best shock-wave path, based on the original image intensity and local pattern orientation. The algorithm was successfully used for validation on numerical Schlieren images, where the shock-wave fluctuation was triggered by vortex shedding. The applicability of the algorithm was finally evaluated in two Schlieren imaging studies: at the trailing edge of supersonic airfoils and for hypersonic research. The program correctly identified the fuzzy flow features present in all applications.
Ultrasonic liver tissue characterization by feature fusion
2012, Expert Systems with Applications
Citation Excerpt :
Data fusion simply combines different domains of raw data to form a new raw data. Decision fusion (Constantinidis, Fairhurst, & Rahman, 2001; Kuncheva & Jain, 2000; Qian et al., 2009) combines multiple classifiers to provide a better result. Feature fusion (Sun, Zeng, Liu, Heng, & Xia, 2005; Yang, Yang, Zhang, & Lu, 2003) selects and combines features extracted from different domains to obtain a better feature subset.
This paper describes a two-stage feature fusion method for ultrasonic liver tissue characterization. The proposed method hierarchically incorporates a genetic-algorithm-based feature selection to automatically select more efficient feature subset to discriminate among ultrasonic images of liver tissue in three states: normal liver, cirrhosis, and hepatoma. Multiple feature spaces are adopted in this paper, including the spatial gray-level dependence matrices (SGLDMs), multiresolution fractal feature vector and multiresolution energy feature vector. Features extracted from different feature spaces may contain complementary information. The feature subsets of different feature spaces are fused and the genetic-algorithm-based feature selection is applied onto the fused feature space to facilitate the two-stage feature fusion. The classification accuracy of the fused feature subset is up to 96.62%. Experimental results demonstrate that the proposed method is capable to select discriminative features among multiple feature vectors to achieve the early detection of hepatoma and cirrhosis based on ultrasonic liver imaging.
Mutual information-based template matching scheme for detection of breast masses: From mammography to digital breast tomosynthesis
2011, Journal of Biomedical Informatics
Development of a computational decision aid for a new medical imaging modality typically is a long and complicated process. It consists of collecting data in the form of images and annotations, development of image processing and pattern recognition algorithms for analysis of the new images and finally testing of the resulting system. Since new imaging modalities are developed more rapidly than ever before, any effort for decreasing the time and cost of this development process could result in maximizing the benefit of the new imaging modality to patients by making the computer aids quickly available to radiologists that interpret the images. In this paper, we make a step in this direction and investigate the possibility of translating the knowledge about the detection problem from one imaging modality to another. Specifically, we present a computer-aided detection (CAD) system for mammographic masses that uses a mutual information-based template matching scheme with intelligently selected templates. We presented principles of template matching with mutual information for mammography before. In this paper, we present an implementation of those principles in a complete computer-aided detection system. The proposed system, through an automatic optimization process, chooses the most useful templates (mammographic regions of interest) using a large database of previously collected and annotated mammograms. Through this process, the knowledge about the task of detecting masses in mammograms is incorporated in the system. Then, we evaluate whether our system developed for screen-film mammograms can be successfully applied not only to other mammograms but also to digital breast tomosynthesis (DBT) reconstructed slices without adding any DBT cases for training. Our rationale is that since mutual information is known to be a robust inter-modality image similarity measure, it has high potential of transferring knowledge between modalities in the context of the mass detection task. Experimental evaluation of the system on mammograms showed competitive performance compared to other mammography CAD systems recently published in the literature. When the system was applied “as-is” to DBT, its performance was notably worse than that for mammograms. However, with a simple additional preprocessing step, the performance of the system reached levels similar to that obtained for mammograms. In conclusion, the presented CAD system not only performed competitively on screen-film mammograms but it also performed robustly on DBT showing that direct transfer of knowledge across breast imaging modalities for mass detection is in fact possible.

View all citing articles on Scopus

About the Author—MICHAEL FAIRHURST has been on the academic staff of the Electronic Engineering Laboratory at the University of Kent since 1972. He has been actively involved in various aspects of research in image analysis and computer vision, with a particular interest in computational architectures for image analysis and the implementation of high performance classification algorithms. Application areas of principal concern include handwritten text reading and document processing, security and biometrics, and medical image analysis.

Professor Fairhurst is a current member and past Chairman of the IEE Professional group E4 on Image Processing and Vision, and has in the past been a member of the Professional Group Committee for Biomedical Engineering. He has been Chairman of several of the IEE International Series of Conferences on Image Processing and Applications, including IPA 99 held in July 1999. He has been a member of many Conference Organising and Programme Committees, including the most recent IWFHR Workshop, and is a member of the British Machine Vision Association. He is a member of the IT and Computing College of the EPSRC. He has published more than 250 papers in the technical literature and has authored an undergraduate textbook on computer vision.

About the Author—A.F.R. RAHMAN graduated in Electronics from Bangladesh University of Engineering and Technology (BUET) in 1992. He subsequently undertook research in pattern recognition, studying the classification of Bengali handwritten characters, for which he was awarded the M.Sc. degree in 1994. He studied for the Ph.D. degree at the University of Kent in the UK and is now on the research staff of the University. Dr. Rahman is a member of the British Machine Vision Association and has published over 40 papers in the technical literature. His research interests are mainly in text recognition, with a special emphasis on the development of multiple expert decision combination structures, but he is also involved in document analysis and neural network processing.

View full text

A new multi-expert decision combination algorithm and its application to the detection of circumscribed masses in digital mammograms

Abstract

Introduction

Section snippets

Overview of operation

Mammography application

Discussion

Summary

Pattern Recognition

A generalised approach to the recognition of structurally similar handwritten characters

IEE Proc. Vision Image Signal Process.

Enhancing multiple expert decision combination strategies through exploitation of a priori information sources

IEE Proc. Vision Image and Signal Process.

A method of combining multiple experts for the recognition of unconstrained handwritten numerals

In IEEE Trans. Pattern Anal. Mach. Intell.

Application of majority voting to pattern recognitionan analysis of its behavior and performance

IEEE Trans. Pattern Anal. Mach. Intell.

A method of combining multiple experts for the recognition of unconstrained handwritten numerals

In IEEE Trans. Pattern Anal. Mach. Intell.

An approach to automated detection of tumours in mammography

IEEE Trans. Med. Imaging