Automatic microcalcification and cluster detection for digital and digitised mammograms

Abstract

In this paper we present a knowledge-based approach for the automatic detection of microcalcifications and clusters in mammographic images. Our proposal is based on using local features extracted from a bank of filters to obtain a local description of the microcalcifications morphology. The developed approach performs an initial training step in order to automatically learn and select the most salient features, which are subsequently used in a boosted classifier to perform the detection of individual microcalcifications. Subsequently, the microcalcification detection method is extended in order to detect clusters. The validity of our approach is extensively demonstrated using two digitised databases and one full-field digital database. The experimental evaluation is performed in terms of ROC analysis for the microcalcification detection and FROC analysis for the cluster detection, resulting in better than 80% sensitivity at 1 false positive cluster per image.

Introduction

Breast calcifications are deposits of calcium inside breast tissue. They appear widespread in the breast and most women will have a few on their mammograms at some point in time, more commonly after menopause [1]. Most calcifications will not be detected during clinical exams or breast self-examination. However, mammography allows to find them long prior to they could move forward into an actual lump. This fact explains why developed countries are adopting the so-called screening programs, which mainly consist in promoting regular women examinations using mammography, usually starting at 40 years and performing them every 2 years.

It is usual to distinguish between two major types of calcifications according to the size: macrocalcifications and microcalcifications. While macrocalcifications are nearly always non-cancerous and need neither additional follow-up nor biopsy, microcalcifications should be diagnosed in more detail. Although about 80% of microcalcifications are typically non-cancerous, when the microcalcifications are new, clustered firmly together, and distributed in specific configurations, they are suspicious signs of breast cancer, most frequently a non-invasive ductal carcinoma in situ. Due to its high spatial resolution, mammography allows to detect microcalcifications at an early stage, a fundamental step for improving prognosis [2], [3]. In a mammogram, microcalcifications appear as small bright spots within an inhomogeneous background. Fig. 1 shows two mammograms from the MIAS database [4] containing a cluster of microcalcifications.

The automatic detection of microcalcifications and clusters is a well-known topic in mammography, as can be seen in the different surveys covering this topic [5], [6]. More recent approaches are by Papadopoulos et al. [7], Pal et al. [8], Rizzi et al. [9] and Yu et al. [10]. Papadopoulos et al. [7] improve previous work [11], which was based on detecting microcalcifications using a neural network, by adding a pre-processing image enhancement step. In their work, different algorithms were tested, obtaining the best results when using the local range modification and the redundant discrete wavelet linear stretching and shrinkage enhancement algorithms. Pal et al. [8] also proposed to use neural networks for microcalcification detection. The first step of their approach consisted in using a multi-layered perceptron network for selecting 29 features that best account for the microcalcification detection from the 87 initially tested. These features are subsequently used to segment the mammograms using another perceptron network. A final step for false positive reduction was necessary for removing thin elongated regions. In this approach, clusters were detected by using a weighted density function which takes the position of the microcalcifications into account. Rizzi et al. [9] proposed a two-stage decomposition wavelet filtering for detecting microcalcifications. The first stage is used to reduce background noise preserving all suspect microcalcifications by thresholding mammograms according to image statistics (mean grey level pixel value and standard deviation), while the second one acts as a hard threshold technique, identifying the microcalcifications from the background. A cluster was considered if more than 3 microcalcifications were detected in a 1 cm² square area. Yu et al. [10] combined model-based and statistical textural features for clustered microcalcifications detection. Firstly, suspicious regions containing microcalcifications were detected using a wavelet filter and two thresholds. Secondly, textural features based on Markov random fields and fractal models together with statistical textural features were extracted from each suspicious region and were classified by a back propagated neural network.

In this paper we present a new approach for the detection of microcalcifications and clusters. The roots of this work can be found in our previous work [12], [13], which was only centred in the detection of individual microcalcifications. In this paper we increase the experimental evaluation of this part, and further, we extend the approach for the detection of clusters, which is more relevant from a clinical point of view. Briefly, the individual microcalcification detection is based on learning the variation in morphology of the microcalcifications using local image features. Afterwards, this set of features is used to train a pixel-based boosting classifier which at each round automatically selects the most salient microcalcification feature. Therefore, when a new mammogram is tested, only the salient features are computed and used to classify each pixel of the mammogram as being part of a microcalcification or actually being normal tissue. Afterwards, the microcalcification clusters are found by inspecting the local neighbourhood of each microcalcification. Note that with our boosting framework we are able to perform both microcalcification and cluster detection without requiring a further classification step as done in previous approaches [7], [8], [9], [10]. Moreover, it is important to remark that we are not dealing with diagnosis in this paper, which is usually performed by means of knowledge-based systems [14], [15], [16].

It is well known that digital mammography allows to improve the detection of microcalcifications thanks to its superior sensitivity [17], and new approaches only dealing with full-field digital microcalcification detection are appearing. Unfortunately, this technology is not yet available in many countries and clinical centres due to its cost. Therefore, reliable automatic approaches able to detect microcalcifications clusters in digitised film plates are still necessary. In the experimental section of this paper, we validate our approach using both technologies. In particular, we used the whole set of 322 mammograms of the MIAS database [4] and a set of 280 mammograms extracted from a non-public full-field digital database. The results show the validity of our approach to deal with mammograms of both natures.

The remainder of this paper is structured as follows. The following section describes the proposed approach for detecting individual microcalcifications in a mammogram. In Section 3 we extend our approach to the detection of microcalcifications clusters. Section 4 presents the experimental set-up designed for testing our approach, while the results are presented in Section 5. The experimental evaluation is done in terms of ROC and FROC analysis. Finally, the paper ends with discussion and conclusions.