Elsevier

Pattern Recognition

Volume 47, Issue 7, July 2014, Pages 2467-2480
Pattern Recognition

Detection of masses and architectural distortions in digital breast tomosynthesis images using fuzzy and a contrario approaches

https://doi.org/10.1016/j.patcog.2014.01.009Get rights and content

Highlights

  • We propose a complete computer aided detection scheme for digital breast tomosynthesis data set.

  • Our method focuses on the detection of architectural distortions and masses.

  • We assess the performance of our system on clinical data and show that it achieves results comparable to state of the art methods.

Abstract

Digital breast tomosynthesis (DBT) is a new 3D imaging technique, which overcomes some limitations of traditional digital mammography. Its development induces an increased amount of data to be processed, thus calling for a computer aided detection system to help the radiologist. Towards this aim, this paper focuses on the detection of masses and architectural distortions in DBT images. A complete detection scheme is proposed, consisting of two parts, called channels, each dedicated to one type of lesions, which are then merged in a final decision step, thus handling correctly the potential overlap between the two types of lesions. The first detection channel exploits the dense kernel nature of masses and the intrinsic imprecision of their attributes in a fuzzy approach. The second detection channel models the convergence characteristics of architectural distortions in an a contrario approach. The experimental results on 101 breasts, including 53 lesions, demonstrate the usefulness of the proposed approach, which leads to a high sensitivity with a reduced number of false positives, and compares favorably to existing approaches.

Introduction

Digital breast tomosynthesis (DBT) is a new three-dimensional (3D) imaging technique aiming at overcoming some limitations of mammography [1]. It has the potential of improving the visibility of breast structures by reducing overlap of tissues. Therefore, the detectability of lesions is potentially increased while false positives (FPs) due to tissues superimposition can be more easily discarded. This comes at the expense of an increased amount of data to be reviewed by the radiologist. Therefore the need for an automatic detection/characterization system, also known as computer aided detection (CAD), dedicated to lesions contained in this new kind of data is increasing, since it may help the radiologist to achieve detection tasks in a reasonable amount of time, while keeping or increasing his sensitivity. Towards this aim, we propose in this paper a complete detection scheme for automated detection of masses and architectural distortions, which are suggestive of malignancy. Fig. 1(a) illustrates the former type of lesions. It is characterized by a dense kernel and ill defined boundaries. In this particular example there are also some spicules. Fig. 1(b) represents the latter, which is characterized by the lack of dense kernel and a strong convergence pattern.

The aim of this paper is not to identify the clinical type of lesions, but detect lesions whatever their type. All lesions we consider have one of the two different appearances in the images. Therefore we design a detection method for each of these broad classes of appearances, regardless of the exact clinical type of the lesion. Since there is no crisp transition between the two appearance classes, the results are finally grouped to provide better overall detection results. While a third type of finding exists (calcification clusters), it is not considered in this paper. Actually, we rely only on commonalities between masses and architectural distortions in order to detect border line findings like strongly spiculated masses. Although this reasoning makes sense for these two types of lesions, it does not for calcification clusters since they do not share characteristics with them. Therefore no gain in overall detection performance can be expected compared to adding an external approach like for instance [2] to our method. For the same reason, CAD systems in the literature are usually dedicated to either masses or calcifications.

Although DBT aims at offering a better characterization of radiological findings, the variability of these structures is still large. Thus, as it is the case for standard mammography, the detection task of pathological patterns remains challenging. Furthermore, DBT being a quite recent imaging technique, the literature does not provide a comprehensive solution to the CAD problem, especially for the case of architectural distortions, which has not been fully addressed so far.

In DBT-CAD, sensitivity can be reported as the ratio of findings that are detected. This can be done either on a DBT volume or on a breast basis. In the first case, each lesion in each DBT volume is considered as a finding to be detected. In the second case, a given lesion contained in a breast that has been acquired from several views (i.e. several DBT volumes) only needs to be detected a single volume. In the literature, the first one is commonly used, mainly because early DBT exams usually contain only one view. However, sometimes both measurements are used [3]. For the sake of clarity, we will always use the first method in this paper when presenting existing results or the performance of our approach. Additionally, specificity will always be given as the number of false positives per DBT-volume.

Preliminary investigations were based on a detection of masses performed within the projections to be used for the reconstruction of the volume [4]. The same authors also proposed to work directly within the DBT volume [5]. While these developments demonstrated the feasibility of a CAD system, specificity remained high (2.2 FP per breast volume at 85% sensitivity).

Other groups worked on both projections and volume based approaches. Reiser et al. [6] proposed to detect masses independently within each projection and to recombine the detection output in the 3D space using visibility angular range of the findings. They also proposed a processing of the volume by a radial gradient filter combined with a maximum intensity projection and a top hat [7]. This processing produces a large number of false positives, therefore it is only suitable as a preliminary detection step of a whole CAD chain.

A fuzzy logic based processing chain has also been proposed [8]. It relies on an independent detection of the suspicious areas within the projections, followed by a segmentation stage using a fuzzy contour framework [9], which allows handling imprecision and uncertainty regarding these steps until the decision making stage using fuzzy decision trees [10]. However, this work mainly focuses on the information aggregation part and lacks a robust initial detection of findings.

More recently a projection and a volume based CAD system were compared by Chan et al. [11]. A hybrid method that combines both approaches was also introduced. The authors used quite a large database (69 malignant lesions) for the evaluation and concluded that their 2D only approach performed worse than the hybrid one. While this work focuses only on mass detection, it will be considered as the reference method in this paper and a comparison with our approach will be provided later on, mainly because of the database size and because it provides the best performance in the literature (1.61 FP per breast volume at 90% sensitivity).

Another kind of approach uses information theory [12], [13] to reduce the number of false positives. While results are promising, the false positive rate still remains high (2.4 FP per breast volume at 90% sensitivity) compared to the method of Chan et al. [11]. This kind of method will be discussed in this paper as well.

Finally, a more recent work has been proposed [3]. The idea is to use a CAD system trained on 2D mammograms on tomosynthesis slabs (aggregation of slices). Using the same evaluation methodology as the other papers, the proposed method gives performance slightly better than the approach of Singh et al. [12]. However, the use of slabs and the poor slice inter-spacing have a negative impact on the localization of the lesion in the volume.

From a technical perspective, designing an efficient CAD system is quite a challenging task mainly because of the difficulty to mathematically define the findings of interest in a way that captures their variability. This results in the difficulty to derive robust operators for their detection/characterization. Actually, existing systems usually rely on two steps. First the detection of potential findings and then false positive reduction. The former should be as sensible as possible in order not to miss cancers. The latter has to drastically reduce the number of false detections. However, if the early detection has a poor specificity, the second stage will likely not reduce it to an acceptable level. In this paper, we introduce a method to cope with all these constraints altogether.

In our work, we propose a detection scheme that addresses both masses and architectural distortion with similar performance to state of the art mass only DBT CAD systems. Since these findings have different appearances and characteristics in the images, we propose an original scheme, composed of two channels, each one being dedicated to one type of lesions. The results provided by these two channels are then aggregated to reach the final decision on the detection. In the first channel, a fuzzy approach is implemented to detect masses. The idea is to model the imprecision on the contours using fuzzy logic, which is suitable to propagate/handle this type of imperfection through the detection process. In the second channel, convergence regions are detected using an a contrario approach. Here the idea is to statistically define the content of a healthy breast in order to detect abnormalities.

An overview of the proposed approach is given in Section 2. Then the detection procedures in the two channels are described in 3 Detection of masses, 4 Detection of lesions without dense kernel. In Section 5 we discuss the aggregation step and the obtained results.

Section snippets

Method overview

In this section we provide a general overview of the proposed approach, and we describe the database used for the experiments and the evaluation.

Detection of masses

The first channel of our detection scheme relies on the detection and classification of dense kernels. In this section, we focus on the description of the detection techniques implemented in the first stage, the segmentation methods and the potential lesions classification.

Detection of lesions without dense kernel

Some of the lesions are not associated with a dense kernel in the image, but rather with a strong convergence pattern. While this kind of findings has partially been addressed in the literature [36], [37] for DBT, a full scheme including robust decision making is still needed. In this section, we present the design and the validation of such a channel, which has previously been introduced in Section 3.

Aggregation and results

In this section, we describe the final aggregation step, detail the experimental results obtained with the proposed complete detection scheme and briefly discuss some implementation details.

Conclusion

In this paper, we have proposed an original method for detecting potential cancerous lesions in DBT images, based on the design of specific methods for masses on one hand, and architectural distortions on the other hand, leading to two detection channels, which are combined in a final decision step. These methods exploit the 3D information provided by the DBT images. Only a few steps are performed on 2D slices, when it is more relevant, and the results are then aggregated in 3D. The results,

Conflict of interest

None declared.

Acknowledgments

This work was partially funded by a grant from ANRT (1165/2006) during the Ph.D. thesis of G. Palma at Telecom ParisTech and GE Healthcare.

Giovanni Palma received the engineering degree in computer science from EPITA, France, in 2005, the master and Ph.D. degrees in signal and image processing from Telecom ParisTech, France, in 2006 and 2010, respectively. His research interests include image processing, mammography, computer aided detection and digital breast tomosynthesis.

References (45)

  • I. Reiser et al.

    Computerized mass detection for digital breast tomosynthesis directly from the projection images

    Med. Phys.

    (2006)
  • I. Reiser, R. Nishikawa, M. Giger, D. Kopans, E. Rafferty, T. Wu, R. Moore, A multi-scale 3D radial gradient filter for...
  • G. Peters, Computer-aided detection for digital breast tomosynthesis (Ph.D. thesis), Ecole Nationale Supérieure des...
  • G. Peters, S. Muller, B. Grosjean, S. Bernard, I. Bloch, A hybrid active contour model for mass detection in digital...
  • G. Palma, G. Peters, S. Muller, I. Bloch, Masses classification using fuzzy active contours and fuzzy decision trees,...
  • H.-P. Chan et al.

    Computer-aided detection of masses in digital tomosynthesis mammography: comparison of three approaches

    Med. Phys.

    (2008)
  • S. Singh et al.

    Automated breast mass detection in 3D reconstructed tomosynthesis volumes: a featureless approach

    Med. Phys.

    (2008)
  • S. Singh, G.D. Tourassi, A.S. Chawla, R.S. Saunders, E. Samei, J.Y. Lo, Computer-aided detection of breast masses in...
  • N. Cristianini et al.

    An Introduction to Support Vector Machines and Other Kernel-based Learning Methods

    (2000)
  • A. Desolneux et al.

    Meaningful alignments

    Int. J. Comput. Vis.

    (2000)
  • G. Palma, I. Bloch, S. Muller, Fuzzy connected filters for fuzzy gray scale images, in: International Conference on...
  • O. Nempont, J. Atif, E. Angelini, I. Bloch, A new fuzzy connectivity class. Application to structural recognition in...
  • Cited by (0)

    Giovanni Palma received the engineering degree in computer science from EPITA, France, in 2005, the master and Ph.D. degrees in signal and image processing from Telecom ParisTech, France, in 2006 and 2010, respectively. His research interests include image processing, mammography, computer aided detection and digital breast tomosynthesis.

    Isabelle Bloch is graduated from the Ecole des Mines de Paris, Paris, France, in 1986, she received the Master׳s degree from the University Paris 12, Paris, in 1987, the Ph.D. degree from the Ecole Nationale Supérieure des Télécommunications (Telecom ParisTech), Paris, in 1990, and the Habilitation degree from the University Paris 5, Paris, in 1995.

    She is currently a Professor with the Signal and Image Processing Department, Telecom ParisTech, in charge of the Image Processing and Understanding Group. Her research interests include 3D image and object processing, computer vision, 3D and fuzzy mathematical morphology, information fusion, fuzzy set theory, structural, graph-based, and knowledge-based object recognition, spatial reasoning, and medical imaging.

    Serge Muller is the chief engineer of breast care business at GE Healthcare. He was graduated from Paris University in 1986 with a Ph.D. degree in spatial technologies and astronomy, and in 2004 in applied mathematics. His current domains of interest are image processing, spectral mammography, digital breast tomosynthesis and robotics.

    View full text