MARCO CONTEXTUAL

grams and Power Law Noise Backgrounds.

Experiment 5 will compare the eectiveness of mirror symmetric and animated display types for the task of detecting a Gaussian blob in paired real mammogram image sections and paired power law noise image pair sections. The experiment will use tightly controlled experimental conditions to ensure that the real mammogram image sections and power law noise image sections are matched on all specications such that the only dierence is the noise type. This will, thus, enable an unmitigated comparison of observer performance with each display type and each noise type.

Chapter 4

Experiment 1: Weak Use of

Symmetry in the Detection of

Simulated Tumours in Paired

Synthetic Mammograms

4.1 Introduction

The introduction to mammography in Chapter 2 took one aspect of the breast radiologist's task and redened it in psychophysical terms. Thus, the task of detecting a tumour or mass from the background of paired mammograms displayed in a mirror symmetric format is presented as the detection of a signal in correlated noise backgrounds displayed in a symmetric format. In the clinical environment, the correlation between the paired mammograms will vary dependent upon the level of normal and abnormal dierences present, and the practice of viewing mammograms back to back in a mirror image display means that, as the level of correlation varies, the level of symmetry presented by the paired images will also vary.

Mirror symmetric presentation is widely recommended as an aid to making abnormalities more salient (Andolina & Lillé, 2010; Bun, 2002; Harvey & March, 2013; Kopans, 2007; Sickles, 2007), and there is an abundance of evidence to support this notion, showing that the human visual system is highly attuned to detect the occurrence of (and by inference, the violation of) visual symmetry (Baylis & Driver, 1994; Julesz, 1971; Koning & Wagemans, 2009; Treder, 2010; Treisman

& Patterson, 1984; Wagemans, 1995).

The eectiveness of mirror symmetric presentation for the detection of large violations, as would be caused by large abnormalities, is not disputed, however, there appears to be no empirical evidence to support the notion that mirror symmetric presentation assists the observer in the detection of a small localised violation, such as may be caused by a small mass or tumour. Indeed, research suggests that mirror symmetric presentation may have limitations for the detection of small localised violations in certain circumstances which may have implications for its use in mammography. First, there is evidence that violations of mirror symmetry may be less easily detected when the violation is remote from the midline of the symmetric display (Barlow & Reeves, 1979; Bruce & Morgan, 1975; Jenkins, 1982). This confers a substantial shortcoming for the use of symmetry in mammography as a localised mass may be present anywhere in the breast, not just close to the midline. Second, previous studies on symmetry detection providing evidence for the eectiveness of mirror symmetric displays as a tool for violation detection have used patterns that are relatively simple (e.g. Baylis & Driver, 2001), with a minimal number of features (e.g. Wenderoth, 1996) or a limited number of violations of the symmetric pattern (e.g. Locher & Wagemans, 1993). A typical pair of mammogram images, however, is not made up of such simple patterns, nor is it likely to be completely symmetric, with normal and abnormal variations in breast tissue potentially reducing the eectiveness of mirror symmetry as an aid to the detection of a localised mass when presented in the traditional side-by-side format.

While not agreeing on the underlying mechanism, several studies have shown the adverse eect of increasingly complex images on symmetry detection, providing explanations of increased levels of information within the images (Tapiovaara, 1990), the increasing density of that information (Rainville & Kingdom, 2002), or the increasing spatial frequency of the patterns within the images (Dakin & Herbert, 1998). In relation to the complexity of paired images, Huang & Pashler (2002), suggested that symmetry detection operates using coarse binary maps that lter individual fea- tures of an image, such as shape, size or colour, that are checked for symmetry (or for violations of symmetry). Huang & Pashler (2002) measured observers' response times to detect symmetric patterns and for all of the features presented found that response times increased as the complexity of the image increased concluding that symmetry detection is spatially inaccurate. When applied to a mammogram image pair, these ndings suggest that symmetry detection would be a very coarse process and would be likely to miss minor violations of the symmetric pattern such as may be caused by a small tumour. This clearly has implications for the eectiveness of symmetry as an aid to the mammographer. Finally, although the occurrence of symmetry in nature is common, it is rarely perfect (Va`rkonyi & Domokos, 2006). This is demonstrated in examples commonly

thought of as symmetric but which rarely are, such as the human face (Lu, 1965) and the snowake which, as Libbrecht (2006, p. 48) noted, "The vast majority show imperfect symmetry, if they show much symmetry at all". The implication here is that the human visual system may have optimally evolved to detect imperfect, natural symmetry but may be insensitive to minor violations of symmetry. This observation was supported by the Tjan & Liu (2005) study which found, using random dot patterns, that the visual system was disproportionately less well attuned to smaller rather than to larger departures from symmetry. The evidence from the study of Tjan & Liu (2005) that that small violations of symmetry are poorly perceived, gives us cause to doubt the eectiveness of such displays for the detection of small localised masses.

The preceding evidence suggests that the value of using symmetric displays to aid in the detec- tion of small masses is, questionable and, whilst this evidence has focused on mirror symmetry, it should be noted that both mirror and repeat symmetric displays are used by the breast radiologist. Thus, the aim of this experiment was to test both mirror and repeat symmetric displays, however, before testing in a clinical scenario, this experiment will establish the theoretical basis by testing the eect of symmetry in a laboratory simulation. Thus, in this experiment we used synthetic images and synthetic tumours and presented two side-by-side noise images to simulate the con- ventional display of two mammograms side-by-side. Observers decided which image contained the synthetic tumour signal. Varying the level of correlation between the background images varies the level of symmetry presented by the images. By varying this during a signal detection task, the experiment is able to determine whether improvements in symmetry lead to improvements in observer performance. The question of whether symmetry helps in the detection of a signal can be analysed theoretically using an ideal observer approach (see Theory section of Experiment 1a on page 125). When attempting to detect a signal known exactly embedded in one of a pair of correlated noise patches, an ideal observer will decorrelate the two patches and cross-correlate a template of the signal with the decorrelated stimuli presented (<signal+noise> or <noise>) (Kay, 1998, p. 106). If the cross-correlation exceeds a criterion level, the observer says "signal present" and "signal absent" otherwise. Decorrelation eectively removes any correlated noise and with a correlation of 1, all the noise will be removed, leaving only the signal (Kay, 1998, p. 111). Thus, an observer who can decorrelate the noise patches will have much better performance than that of an observer who cannot perform such decorrelation. It is interesting to note that the correlation remains the same whether the image pair is presented in a mirror symmetric format or a repeat symmetric format and the ideal observer, therefore, performs the same for both types of symmetry. Whether the human observer can perform decorrelation and use the symmetry of the back- ground is an important question. The ideal observer provides the optimal benchmark against

which other observers can be compared and, even if humans perform poorly in relation to the ideal observer, the comparison provides an insight into the limitations of the human visual system and the requirements of the type of display that would enable the human observer to perform the task optimally. Real observer performance was assessed by measuring the contrast threshold for detecting the signal as a function of the degree of symmetry in the image pairs. The degree of sym- metry was manipulated by varying the correlation between the two noise images. The correlation is dened as:

ρ = σxy

σxσy (4.1)

where σxy is the covariance between pixel intensities in image x and image y, σx is the standard

deviation of image x, and σy is the standard deviation of image y. If the correlation was zero, the

noise was completely unrelated and the two images were completely asymmetric. If the correlation was 1.0, the two images had identical noise (though reected about the vertical axis in the mirror condition) and the image pair had perfect symmetry. Intermediate levels of correlation produced pairs with partial symmetry. If symmetry helps the observer to detect the signal, as suggested by the performance of the ideal observer, the contrast threshold should decline as the correlation increases. In addition, if mirror symmetry as is commonly used in the clinic is helpful for detecting tumours and small masses, we expect that performance will be better for image pairs that have mirror symmetry rather than simply translational symmetry as presented in the repeat condition. Experiment 1a used a Gaussian white noise background to enable a simplied examination of the theoretical basis for the ideal observer. Experiment 1b used a noise background with a 1/f3 _{power spectrum chosen for the similarity of its statistical properties with those of real}

mammogram backgrounds (Burgess et al., 2001). In both experiments the signal to be detected was a Gaussian blob signal, the characteristics of which are similar to the typical mass searched for in real mammograms, as discussed in section 1.6.3 and section 2.15.

The aim of both experiments was to investigate whether mirror and repeat symmetric presen- tations aid the observer in the detection of a signal in correlated noise elds.