1.2. FORMULACIÓN DEL PROBLEMA
2.2.5. Efecto de algunas sustancias peligrosas del agua de consumo humano
If image processing is to be used to assist in the characterisation of microcalcifi cations, they must first be detected, either automatically by the computer or by the radiologist. Obviously automatic detection is preferable. A great deal of research in the processing of medical images has been directed at the problem o f detecting microcalci fications [Chanl988, Fam l988, Davies 1989a, Karssem eijerl992, Shenl993b, Chan 1995b, Q ianl995, Guillemet 1996]. I do not propose to provide a comprehensive review of this work: seven papers included in a recent workshop on digital mammog raphy provide a representative sample of the approaches taken [Brettlel994, C hitrel994, Kegelmey e r1994, Nishikawal994, Parker1994, Strickland 1994, Valatxl994]. The problem is a difficult one, and a variety of different techniques have been applied, most of which share a common structure. First the image is enhanced in some way, to remove noise or to filter out the components o f the signal related to background tissue, secondly there is a detection phase and thirdly the results o f the detection are improved by applying criteria which discriminate between true microcal cifications and false positives. In this section I summarise only the techniques used in the second stage, since techniques of noise reduction are outside the scope of this thesis and the decision criteria employed in the third phase overlap with those described in the next section for the classification of calcifications.
The techniques used to detect microcalcifications fall into two categories, those based on thresholding and those based on matched filters. Thresholding is employed by Chitre et al., Nishikawa et al. and Valatx et al. Chitre uses a threshold on the standard deviation in pixel values in a surrounding region. Nishikawa applies a combination of local and global thresholds on the grey-level of pixels. Valatx applies a local threshold to each image in a stack of images of different resolution.
A variety of filters are used. Brettle et al. use matched filtering in the frequency domain. Kegelmeyer applies a set of filters that have been used by other investigators and uses the results as input to a binary decision tree. Parker describes a filter based on the idea of hysteresis: line profiles are taken through the image at different orientations and then analysed by a filter which is analogous to passing a loop of wire over a one dimensional signal. The loop will detect all local maxima where the difference in grey- level is larger than the loop. Strickland et al. use a ‘wavelet’ transform which acts as a bank o f multiscale matched filters.
There is no easy way to compare the performance o f these different techniques on the basis of the published articles. The results are based on different sets of data, the criteria used to define successful detection vary and the results are reported in quite different ways. Since the detection of calcifications is not o f intrinsic interest for this thesis, I decided not to attempt to identify the best detection technique but to re implement the filters described by Kegelmeyer and to test whether these could form the basis of a detection algorithm which would allow the application o f measures for the classification of calcifications. I decided to take advantage of Kegelmeyer’s work since this in turn built on techniques developed by other authors and therefore seemed likely to be generally applicable. Details of the implementation are provided in Chapter Seven.
6.5
Image Processing for the Classification of Calcifications
The aim of this work is to provide a decision aid for the classification of calcifi cations which is able to provide information based not only on knowledge-based reasoning but also on data derived from image processing. Nine sets of authors have recently published papers on the use of various kinds of computer system to help distinguish benign from malignant microcalcifications. In this section I summarise the
approaches taken, and then consider how elements of this work could best be incorpo rated into a décision-support tool for image interpretation.
In the following nine sub-sections I give outline summaries of these papers. There is a common structure to the approaches taken, and this determines the form of the outline summaries. First, each system has some mechanism for detecting calcifi cation data in images, looking either for individual microcalcifications, clusters or regions of microcalcifications. Different measures are then taken of the data and these measures are used as the basis for some form of classification rule. Different techniques are used to generate the classification rule and in some systems a subset of useful measures is first identified. In the next nine sub-sections, the aim of the research is stated as well as the method used to identify the calcifications and their clusters, the set of measures considered, the set of measures finally adopted - or described as having been most useful - the technique used to identify the classification, and the results of any test of the classification. Detailed definitions are provided in the Appendices.
6.5.1
Magnin [Magninl989]
Aim
The aim was to distinguish two types of microcalcification: lobular and intra ductal. The value of this distinction is that lobular calcifications are more likely to be benign. The paper deals with the properties of individual microcalcifications and not those o f clusters.
D etection o f microcalcifications
The microcalcifications were identified automatically from a pre-processed image using a combination of thresholds, one computed by adaptive grey-level thresh olding (adaptive in the sense that the threshold is computed for each pixel on the basis of statistics in a local region around the pixel) and the other based on a residual energy
measure (the sum of the squared differences between pixel values and a threshold for the region).
M easures considered
Seven measures were considered: perimeter area compactness eccentricity convexity lengthening rate mean ray vector
Classifier
No classifier was developed.
M easures used
No formal analysis was done but it is suggested that four measures can be used to distinguish the two types of microcalcification:
• compactness • convexity • lengthening rate • mean ray vector
1. Compactness is a measure of the degree to which a shape resembles a circle; its definition is given in the