Classification images of two right hemisphere patients: A window into the attentional mechanisms of spatial neglect

Transcription

1 available at Research Report Classification images of two right hemisphere patients: A window into the attentional mechanisms of spatial neglect Steven Shimozaki a,, Alan Kingstone b, Bettina Olk c, Robert Stowe d, Miguel Eckstein a a Department of Psychology, University of California, Santa Barbara, CA 93106, USA b Department of Psychology, University of British Columbia, Vancouver, British Columbia, Canada c School of Humanities and Social Sciences, International University Bremen, Bremen, Germany d Departments of Psychiatry and Neurology, University of British Columbia School of Medicine and Riverview Hospital, Canada ARTICLE INFO ABSTRACT Article history: Accepted 9 January 2006 Available online 21 February 2006 Keywords: Visual attention Spatial neglect Cueing Classification image Bayesian observer While spatial neglect most commonly occurs after right hemisphere lesions, damage to diverse areas within the right hemisphere may lead to neglect, possibly through different mechanisms. To identify potentially different causes of neglect, the visual information used (the perceptual template ) in a ing task was estimated with a novel technique known as classification images for five normal observers and two male patients with righthemisphere lesions and previous histories of spatial neglect (CM, age 85; HL, age 69). Observers made a yes/no decision on the presence of a White X checkerboard signal (1.5 ) at one of two locations, with trial-to-trial stimulus noise added to the 9 checkerboard squares. Prior to the stimulus, a peripheral pre (140 ms) indicated the signal location with 80% validity. The ing effects and estimated perceptual templates for the normal observers showed no visual field differences. Consistent with previous studies of spatial neglect, both patients had difficulty with left (contralesional) signals when preceded by a right (ipsilesional). Despite similar behavioral results, the patients' estimated perceptual templates in the left field suggested two different types of attentional deficits. For CM, the left template matched the signal with left-sided s but was opposite in sign to the signal with right-sided s, suggesting a severely disrupted selective attentional strategy. For HL, the left templates indicated a general uncertainty in localizing the signal regardless of the 's field. In conclusion, the classification images suggested different underlying mechanisms of neglect for these two patients with similar behavioral results and hold promise in further elucidating the underlying attentional mechanisms of spatial neglect Elsevier B.V. All rights reserved. 1. Introduction 1.1. Spatial neglect and ing Spatial neglect can occur after a unilateral brain injury and describes a syndrome in which the patient ignores visual information in the hemifield opposite to the side of the lesion (for a review, see Rafal, 1994; Lezak, 1995; Gazzaniga, 1998). It is not a purely visual deficit as it can be dissociated from visual field loss (hemianopia), and it is commonly assumed that hemineglect reflects an attentional deficit (although many have argued that hemineglect also reflects a deficit in the Corresponding author. Fax: address: shimozak@psych.ucsb.edu (S. Shimozaki). URL: (S. Shimozaki) /$ see front matter 2006 Elsevier B.V. All rights reserved. doi: /j.brainres

2 27 representation of space, see Bisiach, 1993; Bisiach and Luzzatti, 1978). Aside from its clinical implications, many cognitive neuroscientists view neglect as an excellent model for examining the brain mechanisms that mediate attentional orienting. While neglect may result from a brain injury to one of many regions, the most common and most severe cases of hemineglect result from brain injuries to the right parietal lobe. This site coincides with a number of primate single-cell recording studies (Mountcastle et al., 1975; Robinson et al., 1995; Colby et al., 1995, 1996; Andersen, 1995) and brain imaging studies on normal human observers (Darby et al., 1996; Corbetta et al., 1978; Courtney et al., 1996) suggesting that a primary responsibility of the parietal cortex is the representation and integration of visuo-spatial information. Typically, neglect is most severe at the onset of injury, and, over the course of a year, the patient recovers to a chronic level of deficit (which can approach normal levels of functionality). A significant paradigm in the general study of attention is known as the ing task (Posner, 1980). In this task, observers detect a target stimulus that could appear at one of two or more locations. Before the stimulus appears, a pre appears that indicates the probable location of the forthcoming target. Trials in which the target appears at the d location are known as valid trials, while those in which the target appears at an und location are known as invalid trials. There may also be a neutral trial type, in which no reliable information is given about where a target may appear. The typical finding for these different pre conditions is that with normal observers performance is fastest and/or most accurate when a target appears at a validly d location, worst when a target appears at an invalid (und) location, and intermediate for neutral trials. Aside from studies on normal observers, the ing task has been applied in both the diagnosis and study of neglect. In a now classic investigation, Posner et al. (1984) found that hemineglect patients were severely impaired when the pre appeared in the ipsilesional (good) visual field and the target appeared in the contralesional (bad) visual field. Thus, hemineglect patients, unlike normal observers, manifest a significantly larger ing effect when the pre appears in the ipsilesional visual field than when it appears at the contralesional visual field. From this study and subsequent investigations like it, Posner and colleagues developed a theory of attention in which the parietal cortex is responsible for disengaging attention from a selected spatial location. Hence, when the parietal cortex is lesioned, as it is for most neglect patients, performance is severely compromised when attention must be disengaged from an ipsilesional and shifted to a contralesional target. While this disengage deficit provides an accurate qualitative characterization of the attentional deficit that characterizes neglect patients, it fails to provide a quantitative description of the deficit. The aim of the present study is to begin to address this shortcoming Description of ing task Fig. 1 depicts the ing task used in the current study. Observers performed a yes/no contrast discrimination of a 3 3 checkerboard pattern configured as a white X. On half the trials, the signal appeared for 40 ms (normals) or 140 ms (patients) at either one of two locations (left and right), with a 140 ms pre indicating the probable location of the signal with 80% validity. Observers had to Fig. 1 Trial types in the d yes/no contrast discrimination task, right-sided s only. Observers judged upon the presence of a high contrast checkerboard white X (140 ms, either 2.5 left or right of central fixation) appearing on half the trials. A pre (2.5 square, 140 ms) indicated the signal location with 80% validity on signal present trials.

3 28 BRAIN RESEARCH 1080 (2006) judge if the signal appeared (and not where) during a particular trial. Thus, there were three trial types, valid signal present (upper), invalid signal present (center), and signal absent (lower), as depicted in Fig. 1. One might view this task as the equivalent to the original ing task performed by Posner (1980), except measuring accuracy as opposed to reaction time. As depicted in Fig. 2, the luminances of each of the 9 squares comprising the checkerboards were not constant but had a different sample of random noise added for each trial. This was necessary to compute the classification images described later. The classification image technique developed by Ahumada and Lovell (1971) gives a direct estimate of the perceptual filter (template) of an observer based on how an observer's judgments correspond to the image noise that is added to the stimulus A quantitative model of the ing task Recently, investigators have applied quantitative tools and computational modeling that have been successfully used to study basic visual phenomena to investigate the mechanisms mediating visual attention (see Dosher and Lu, 2000a,b; Eckstein et al., 1997). For the Posner ing task, one starting point within this general approach is to consider the theoretical framework of the ideal Bayesian observer (Eckstein et al., 2002; Shimozaki et al., 2003, see Kersten et al., 2004 for a review). Ideal observer analysis determines the algorithm that achieves the best possible performance for a given task, and it is a potentially powerful technique to analyze performance in visual tasks generally (Barlow, 1978; Burgess et al., 1981; Kersten, 1984; Pelli, 1985; Liu et al., 1995; Tjan et al., 1995; Bennett et al., 1999; Gold et al., 1999) and visual attention tasks specifically (Eckstein et al., 2002; Shimozaki et al., 2003). First, ideal observers can be used as an absolute standard of performance that can be compared to human performance. Second, ideal observers can also serve as a starting point for quantitative modeling. As real observers, humans cannot match the performance of the ideal observer, and the difference between human and ideal observers can be modeled with respect to the ideal observer as various sources of inefficiency of human perceptual and decision processing. Fig. 3 presents a quantitative model of the Posner ing task based on the ideal observer known as the Weighted Likelihood model (Eckstein et al., 2002; Shimozaki et al., 2003; also see Appendix B). The responses of the Weighted Likelihood model (x c and x uc ) are based on the match (correlation) between a template at each location and the stimuli at each of the two possible signal locations. From these responses, the model calculates the probability of signal presence at each location (the likelihoods). These likelihoods are then multiplied by separate scalar values, or weighted, with the weight for the d location (w c ) typically being larger than that for the und location (w uc ). Finally, the two weighted likelihoods are added together and compared to a criterion to make a yes or no decision upon signal presence. The Weighted Likelihood model is equivalent to the ideal observer under certain conditions (see Appendix B), and we will focus on two of these conditions in this study. The first primary condition is that the weights (w c and w uc ) are equal to the validities (as the validities correspond to the prior probabilities of the signal presence, given the location). We can assess this condition by comparing the ideal weights (which are w c = 0.80 and w uc = 0.20 in this experiment) to the best estimate of the human observers' weights. Note that, for the ideal observer, the differential weighting is the only mechanism of attention, thus the ideal observer is sometimes called a purely selective model of attention. The second primary condition is that the templates at each location are an ideal shape, which in this task is simply the checkerboard signal. We assessed the template shapes in this study with the classification image technique. As mentioned earlier, human observers (both normal and patient) may be modeled within the framework of an ideal observer as losses of efficiency compared to the ideal observer. For the Weighted Likelihood model of the ing task, these include: inadequate weighting in the integration of information from d and und location (the first condition above), templates that do not match the signal (the second condition above), noise internal to the observer, and uncertainty about the signal location leading to multiple templates at different locations instead of one template (Barlow, 1978; Burgess et al., 1981; Pelli, 1985; Eckstein et al., 2002; Shimozaki et al., 2003) Basics of the classification image technique Fig. 2 The signal with and without noise. On each trial, a random luminance value was added independently to each square of the checkerboard (represented on the right); this was necessary to compute the classification images. For visibility, the signal in the figure has a higher contrast than that used in the experiment. Recently, we have assessed a quantitative distinction between these mechanisms using a technique known as classification images (Eckstein et al., 2002). The classification image technique combines the noise samples in the visual stimuli across particular trial outcomes (such as false alarms in a yes/ no task) to estimate the spatial profile of the information used in a visual task, or roughly speaking, the shape of the underlying perceptual templates (Ahumada and Lovell, 1971). It has been used successfully to study a number of different visual phenomena, including Vernier acuity (Ahumada, 1996; Beard and Ahumada, 1998) and other position discrimination judgments (Levi and Klein, 2002; Li et al., 2004), depth (Neri et al., 1999), and Kaniza squares and other illusory contours (Gold et al., 2000).

4 29 Fig. 3 Schematic of the Weighted Likelihood model in the ing task. The model response of a valid signal present trial on the right is depicted. See Appendix B for a description of the model. Furthermore, we have used classification images previously to assess visual attention in normal observers (Eckstein et al., 2002, 2004) and have found that the technique can be used to test predictions of different types of attentional mechanisms that are otherwise indistinguishable by measuring behavioral performance. Here, we propose to use the classification image technique to identify the mechanisms mediating the attentional loss in hemineglect patients as assessed by the ing paradigm. Specifically, we assess the attentional disruption mediating the larger ing effect when the appears at the ipsilesional visual field. Fig. 4 illustrates the calculation of the classification image for the simple ing task used in this study. Assume that we have selected those trials in which the only appeared on the right (upper box). Also assume that the second line represents a sequence of stimuli for a number of trials in the basic ing task, with each trial presented with an independent sample of image noise (for presentation purposes, the samples of noise are not shown). In the figure, a hypothetical observer's response to the trial appears under each stimulus. The first step in calculating classification images is selecting individual trials by the type of trial, defined by both stimulus presence and the observer's response. For example, the figure depicts choosing those trials in which no signal appeared, but the observer made an error and responded yes (a false alarm ). Then, all the noise fields leading to a false alarm decision are averaged to create the classification image. Intuitively, an error in a false alarm trial indicates that some part of the noise field (which varies from trial to trial) led the observer to believe the signal was present; in other words, the image noise on that trial mimicked an aspect of the pattern that the observer considers to be the signal. Thus, the classification image (for false alarms) is an average of the spatial information used by an observer to judge signal Fig. 4 Calculating classification images for the validity task. The figure represents a series of 7 trials and responses in which the appears on the right for a hypothetical observer. The stimuli are presented without noise for clarity but would be presented with added luminance noise in an actual experiment (see Fig. 2). The classification image is the mean of all the noise fields resulting in a false alarm (a signal present response to a signal absent stimulus). Two classification images may be calculated, one for the d location and one for the und location; also, there are two classification images (d and und) for trials in which the appears on the left.

5 30 BRAIN RESEARCH 1080 (2006) presence, or, in other words, the perceptual template of the observer. It should be noted that there are two separate classification images in Fig. 4, one for the valid-d location (in this case, the right ipsilesional field) and one for the invalid-und location (the left contralesional field). We may also calculate classification images for trials in which the appeared on the left for the two locations. Finally, it should be noted that classification images can be calculated from the other response outcomes in this task other than the False Alarms (Correct Rejection, Valid Hit, Valid Miss, Invalid Hit, Invalid Miss) and that the classification images across these different outcomes may be combined into more general classification images (Beard and Ahumada, 1998; Ahumada, 2002; Murray et al., 2002). If we consider the classification images across the different outcomes, a final division is whether the checkerboard signal appeared at the location on that trial (signal) or not (no-signal). The no-signal classification images would include trials in which the signal appeared in the other location and trials in which the signal did not appear at all. Thus, for each location (left and right), we may define six classification images: d, further divided into valid dsignal and valid d-no signal, and invalid und, further divided into an invalid und-signal and invalid und-no signal Possible classification images for different attentional disruption mechanisms in neglect patients Fig. 5 gives some hypothetical classification images, the left halves of the figures giving the 2-dimensional classification images and the right halves of the figures give the same classification images presented as one-dimensional plots. The plots present the nine values of the classification images in serial reading order, starting at the upper left and going left to right and top to down. Fig. 5A depicts the checkerboard signal itself, which, in the context of this experiment, is also the best possible template to maximize performance; intuitively, this template uses all the information available and thus corresponds to the ideal observer template. Figs. 5B through F depict possible classification images for some different types of attentional deficits. In the case of right hemisphere damage, these classification images would be expected for the location in the left visual field. Fig. 5B depicts the case in which no classification image is found or zero values are found for the entire checkerboard. Such a classification image indicates that the observer did not or could not use any of the information presented at that location. Fig. 5C depicts a classification image in which the observer uses only the central square of the checkerboard. Fig. 5D presents the special case of object neglect, in which the observer does not use information from the left side of the checkerboard (e.g., Behrmann and Tipper, 1994; Tipper and Behrmann, 1996). In this special case, potentially, this pattern of neglect might be expected from either the left or right visual field. In the cases represented in Figs. 5B through D, the observer does not use all or part of the information available. Previously, we have characterized this type of deficit as the observer having a template or filter of the incorrect shape (as compared to an ideal observer). The last two cases presented in Figs. 5E and F represent a different type of deficit, one in which the observer uses a template of the correct shape, but of the wrong value or amplitude. The ideal observer can be described as having two components, a template of the correct shape and a correct strategy of the use of the information in the final decision. This strategy is expressed as the weight given to the information at each location, which is represented as the amplitude of the template (relative to the amplitude of the template at the other location). Roughly, it represents how much importance to give the information at each location. Fig. 5E represents a case in which not enough weight is given to the information at the d location. 1 Fig. 5F represents an extreme case in which an opposite weighting is given to the information at the d location. This last case describes a situation in which the observer takes information in favor of target presence as evidence against target presence. We assessed the behavioral performance (accuracy) and classification images of 5 normal and 2 patient observers in the ing task described above. Both patients had suffered a right hemisphere stroke at least 1 year prior to this study; these injuries resulted in neglect for both patients. At the time of this study, assessment by standard neuropsychological examination indicated that both patients' neglect had improved since their injuries, with one patient testing as fully recovered. Despite the neuropsychological results, this study found clear attentional deficits for both patients. In addition, even though the two patients had similar behavioral effects, the classification images clearly indicate that different underlying attentional mechanisms are responsible for the patients' performances. 2. Results and discussion Fig. 6 depicts a schematic of all conditions in the simple ing task, divided by location and visual field, with each cell depicting only the pre display and the stimulus display. Table 1 presents the hit and false alarm rates for the observers, again divided by location and visual field. Note that signals on the left were valid trials for left s and invalid trials for right s, and vice versa for signals on the right. For both Table 1 (hit and false alarm rates) and Table 2 (ing effects), the means and standard errors for each measure are computed over the values calculated for each individual session. All observers were considerably above chance in this task (chance performance would be indicated by equal hit and false alarm rates), but they were not close to perfect performance (hit rates at 100% and false alarm rates at 0%). We defined ing effects for this task as the difference between the hit rate for validly d signals minus the hit rate for invalidly d signals (Eckstein et al., 2002; Shimozaki et al., 2003). The ing effects were also divided by side, so that the ing effect for left s was defined as hit rate (cell 1 It should be noted that a lower amplitude classification image in this case could have two causes. The first is an underweighting of the likelihoods, expressed as w c and w uc in Fig. 3. The second is that the classification image might be subject to more noise inherent to the observer (internal noise), as noted by Ahumada (2002).

6 31 Fig. 5 Hypothesized classification images. For each figure, the left shows the 2-dimensional classification image and the right shows a 1-dimensional representation of the same classification image, in which each of the 9 squares of the checkerboard are represented in reading order, going from left to right and from up to down, starting at the upper left. (A) The ideal observer, which is the signal itself. The ideal observer is represented as the dashed line in parts B through F. (B) No classification image (suggesting that the observer does not use any information from that location). (C) The observer uses only the center point (center only). (D) The observer cannot use information from the left side of the stimulus (local neglect). (E) The observer has the ideal observer template shape but does not give the information from that location enough weight (under-weighting). (F) The observer has the ideal observer template shape but gives the information from that location a weight opposite to the appropriate weighting (opposite weighting). a) the hit rate (cell b), and the ing effect for right s was defined as hit rate (cell e) the hit rate (cell d). Table 2 and Fig. 7 present the ing effects by side for all observers (Column 1 ing effect, Column 2 standard error across sessions). Columns 3, 4, and 5 present the results of tests of significant (nonzero) ing effects, assessed by single-sample t statistics. Columns 6, 7, and 8 present results of the fits of the observer results to the Weighted Likelihood model (see Appendix B). The three parameters from the fits were the weight for the d location (w c ), a measure of sensitivity, or difficulty of the task for the observer (d ), and log(crit), a measure of decision bias. The ideal observer has a weight equal to the validity (0.80); therefore, humans may be assessed by comparing their fitted weights to the ideal weight. This assessment is not absolute performance (which would be percent correct or d ), but rather the use of the information provided by the and appropriate use of the combination of information across the d and und locations. Two observers showed no (PC left s, RP) or negative (PC, right s) ing effects, indicating a general lack of the use of the pre in the task. We (Shimozaki and Eckstein, 2003) have found that normal observers (especially naive observers, such as PC and RP) can vary in the size of their ing effects in these types of ing tasks. The results for the other three normal observers generally indicated modest overall ing effects, with the exception of EM having a near-zero ing effect for right-sided s. Furthermore, the fits to the Weighted Likelihood model showed that all ing effects for the normal observers were less than or equal to the ing effects expected from the ideal observer (as indicated by all weights being less than or equal to 0.80). Finally, in general, the normal observers also showed a tendency for slightly larger ing effects for left-sided s. Analyses by two-sample t tests on left- and right-sided ing effects sampled by each individual session found that this

7 32 BRAIN RESEARCH 1080 (2006) Fig. 6 Schematic of all trial types in the ing task, divided by visual field, with each cell depicting only the pre display and the stimulus display. tendency was not significant for the normal observers, except for PC (t(18) = 3.548, P = ). From previous research (Posner et al., 1984), neglect patients should suffer most severely when detecting a signal in the contralesional left field after receiving an invalid ipsilesional right (Fig. 1, invalid signal present trial; Fig. 6, cell d). Therefore, for the neglect patients, it is anticipated that the ing effect with ipsilesional (right) s (hit rate (cell e) hit rate (cell d)) will be large. Conversely, the ing effect with contralesional (left) s (hit rate (cell a) hit rate (cell b)) should be comparable to normal observers. As expected, the patients did have a large ing effect with right-sided s, and the patients' ing effects for left-sided s were smaller and comparable to the normal observers' ing effects. Analyses on these ing effects by twosample t test analyses, sampled by calculating the ing effects for each individual session, found a significant difference between left- and right-sided s for both patients (HL, t(27) = 2.822, P = ; CM, t(55) = 2.184, P = ), unlike the normal observers. Furthermore, the fits for the patients with right-sided s to the Weighted Likelihood model indicated a greater weighting of the d information than the ideal observer (HL = 0.98, CM = 0.97), while the same fits for left-sided s indicated that weights were slightly less than the ideal observer (HL = 0.72; CM = 0.68). While some of the normal observers were not optimal (with two normal observers having no or negative ing effects), the suboptimal pattern of greater weighting of d information shown by the patients (with right-sided s) was not seen with any of the normal observers. Additionally, the patients' larger ing effect for right-sided s is opposite the general tendency of the normal observers, who generally had slightly larger ing effects with left-sided s. Finally, the similar size of ing effects for the patients with left-sided s to the normal observers (for both left- and right-sided s) suggests that the comparison between the patients and normal observers in this study was reasonable. Figs depict the normal observers' classification images, and Figs. 13 and 14 depict the patients' classification images, calculated from about 2500 trials for the normal observers and 820 (HL) and 1641 (CM) total trials for the patients. As mentioned earlier, classification images were separated by location (left or right), presence, and signal presence. Each square of the checkerboard pattern is presented along the x axis in reading order, top left to bottom right. The ideal observer template for this task (the signal itself, Fig. 5A) is represented in Figs as the alternating dashed line, consistent with the alternating white and black squares. For these figures, the amplitudes of the ideal observer template were chosen by the best χ 2 (chi-square) fit to the human observers' templates (solid lines). We also assessed the shape of the classification images by measuring the correlations of the classification images with the ideal observer template (the signal). Table 3a gives the correlations of the normal observers' classification images with the ideal observer template, and Table 3b gives the correlations of the patients' classification images with the ideal observer template. As expected, in general, the normal observers' estimated perceptual template shapes were well-matched to the ideal observer's template (the signal) across the types of classification images and showed no hemispheric differences. The one exception is in the und-no signal classification images across the normal observers. There were near-zero und-no signal classification images on both sides for TB, and on the right side for EM, leading to low correlations with the ideal observer for these classification images. The other observers also showed a decrease in the amplitude of the und-no signal classification images, but not to the degree of TB and EM. The classification images of both patients indicated an effect in the (left) contralesional field. Furthermore, despite the similarity of the behavioral (ing) results between the two patients, the pattern of effects shown by the classification images differed for the two patients. For HL, the classification

8 33 Table 1 Hit and false alarm rates for the observers by side Signal on left Signal on right Signal absent Hit rate SE hit rate Hit rate SE hit rate FA rate SE FA rate Normal observers EM Left Right PC Left Right RP Left Right SS Left Right TB Left Right Patients HL Left Right CM Left Right Hits = correct signal present response in signal present trial. False alarm (FA) = incorrect signal present response in signal absent trial. Error bars are standard errors of ing effects, over hit and false alarm rates calculated for each individual session. images in the left hemifield were uncorrelated with the ideal observer when the signal was not present. On the other hand, regardless of whether the left or right side was d, the classification images were positively correlated with the ideal observer when the signal was present. This pattern was indicated by a significant difference between the left signal and no-signal classification images, collapsed over presence (2-sample Hotelling T 2 = , F(9,386) = 2.013, P = ). Ahumada (2002) found that this pattern of results is consistent with a specific difficulty for stimuli with phase information, such as grating patterns or the checkerboard, such that the observer was unable to determine the exact location of the potential signal. This type of difficulty is called location uncertainty (Pelli, 1985) and may be modeled within the framework of models like the Weighted Likelihood model as the placement of multiple templates (instead of the one template) in slightly different locations. In this framework, when the signal is present, only the template matched to the signal location responds well, leading to a classification image also well-matched to the signal. However, when the signal is absent, the use of the multiple templates leads to a classification image that loses (or blurs ) the phase properties of the signal, giving a flat classification image. For CM in the left (contralesional) field, the classification images were positively correlated with the signal, except when there was neither a signal nor a at the ipsilesional location. In this last case, the classification image was negatively correlated with the signal, suggesting that CM took evidence consistent with signal presence as evidence against signal presence. In agreement with this pattern, there was a significant difference between the left d vs. und nosignal classification images (2-sample Hotelling T 2 = , F (9,1160) = 2.237, P = ). Clearly, this represents a severely suboptimal attentional strategy for CM, and it also suggests a different type of attentional loss than HL. Consistent with this suggested difference between HL and CM, significant differences were found between HL and CM for both the left d no-signal classification images (2-sample Hotelling T 2 = , F(9,718) = 1.960, P = ) and the left und no-signal classification images (2-sample Hotelling T 2 = , F (9,1035) = 1.923, P = ). Aside from the contralesional (left) classification images, the classification images in the ipsilesional (right) field for the two patients also suggested a potential difficulty in using information in the right field. For HL, no classification image was found in the und trials overall, and, for CM, no classification image was found in the und-no signal condition. These results for the ipsilesional field seem to be consistent with the suggestion of several authors (Làdavas et al., 1990; Olk et al., 2002; Rusconi et al., 2002; Wright et al., 2005) that the neglect syndrome is not restricted to hemifield and may instead reflect a more general graded spatial deficit across hemifields. The last analyses assessed the relationship of the observed classification images with the observed behavioral results. In other words, can the classification images be used to predict the hit and false alarm rates of the observers? As the observers had no a priori information regarding signal presence, we restricted our analysis to the classification images collapsed over signal presence/absence. These classification images may be found at the tops of Figs and denoted as All. The analyses of ing effects for left-sided s included the Cued Left (top left, Figs. 8 14A) classification images and the Und Right (top right, Figs. 8 14B) classification images, and the analyses of ing effects for right-sided s included the Cued Right (top right, Figs. 8 14A) classification images and the Und Left (top left, Figs. 8 14B) classification images. In the remaining figures, analyses for left-sided s using the Cued Left and Und Right classification images are denoted Cued L Unc R and analyses for right-sided s using the Cued Right and Und Left classification images are denoted Cued L Unc R. Fig. 15 depicts the correlations of the Cued Left/Und Right and the Cued Right/Und Left classification images with the ideal observer template (the signal itself); these values are also presented in Table 3 under Cued and Und. We may use these correlations as an assessment of how well (or how efficient) the human observer could potentially use the information at the d and und locations, with high positive correlations indicating good matches to the ideal observer. Fig. 16 depicts the correlations of the Cued Left with

9 34 BRAIN RESEARCH 1080 (2006) Table 2 Cueing effects for the observers by side Cue Cueing SE T df P value w c d log(crit) EM Left Right PC Left Right RP Left Right SS Left Right TB Left Right Patients HL Left Right CM Left Right Column 1: ing effect = hit rate for valid signal location hit rate for invalid signal location for all s on the same side. With respect to Fig. 6: left side s, ing = hit rate for cell e hit rate for cell d; right side s, ing = hit rate for cell a hit rate for cell b. Column 2: standard errors of ing effects, from ing effects calculated for each individual session. Columns 3, 4, and 5: tests of significant (nonzero) ing effects, single-sample t tests, degrees of freedom, and P values. Columns 6, 7, and 8: results of fits to the Weighted Likelihood model, w c = weight for d location, d = measure of sensitivity/difficulty, crit = decision criterion. See Appendix B for details. the Und Right classification images, and the correlations of the Cued Right with the Und Left classification images. These correlations indicate the similarity of the potential use of information at the d and und locations. Considering the correlations in both, Figs. 15 and 16 suggest two distinct patterns among the normal observers. The d and und classification images of PC, RP, and SS have uniformly high correlations with the ideal observer (Fig. 15) and each other (Fig. 16). Thus, these three observers may be characterized as having nearly ideal templates at both the d and und locations for both left- and right-sided s. For EM and TB, Fig. 15 indicates that the d classification images had high correlations with the ideal observer, while the und classification images had lower correlations with the ideal observer on the right for EM and on both sides for TB. Fig. 16 also shows that the Cued/opposite Und correlations were correspondingly low in these conditions for these two observers. The pattern of correlations for the patients shown in Figs. 15 and 16 indicates distinct differences from the normal observers and from each other. For HL, Fig. 15 shows that the Cued Right classification image correlation with the ideal observer was high and similar to the normal observers', Fig. 7 Cueing effects for all observers divided by side, measured as valid hit rate invalid hit rate. The error bars are standard errors over ing effects calculated for each individual session. * significant difference between ing effects for left-sided vs. right-sided s (P b 0.05).

10 35 Fig. 8 Classification images for EM, a normal observer. The classification images are presented as 1-dimensional plots, with each of the 9 squares of the checkerboard represented in reading order, going from left to right and from up to down, starting at the upper left. The signal (ideal observer) is represented as the dashed line, with the amplitudes matched to the best chi-square fit to the unsigned amplitudes of the human observers' templates (solid lines). The symbols correspond to the (squares), signal (X), and classification image (shaded region) locations relative to the fixation point for trials used for each classification image. The cell letters correspond to the cells in Fig. 6. (A) Cued ( appeared on same side), d-signal, d-no signal. (B) Und ( appeared on same side), und-signal, und-no signal. and his Cued Left classification image correlation with the ideal observer was positive but somewhat lower than the Cued Right correlation. The und classification image correlations with the ideal observer, however, were near-zero for both sides. As expected from these correlations with the ideal observer, Fig. 16 shows that both Cued/opposite Und correlations were near zero for HL. For CM, Fig. 15 indicates that his d classification image correlations with the ideal observer on both sides were high and similar to the normal observers. The Und Right correlation with the ideal observer was low, while the Und Left correlation with the ideal observer was negative (indicating the opposite weighting of information by CM). Correspondingly, Fig. 16 indicates that CM's Cued Left/Und Right correlation was low and that his Cued Right/Und Left correlation was negative. Tables 4 6 and Figs summarize the results of simulations of three models to predict human performance from the classification images. Cueing effects may be generated within the constraints of the Weighted Likelihood model in two ways (excluding internal noise added to the und location 1 ). The first assumes that that there is a differential weighting of the d and und locations, irrespective of the templates at the d and und locations. The ideal observer exhibits a ing effect in this way as it weights the likelihoods by validity and has the same (optimal) template at the d and und locations (Eckstein et al., 2002; Shimozaki et al., 2003). The second way is a difference in the template shapes such that the und template is a poorer match to the ideal observer than the d template, irrespective of the weights. The goal of the simulations was to assess the relative importance of weighting changes and any shape differences of the d and und classification images in predicting the observers' results. The models were fit to the valid and invalid hit rates and the false alarm rates of the observer, and these results are presented in Tables 4 6. However, the fits to the false alarm rates were generally quite good, so that the overall fits may be described as a fit to the ing effect (valid hit rate invalid hit rate), which are presented in Figs Table 4 and Fig. 17 give the results of the simulations using the normalized classification images and equal weighting (w c = w uc = 0.50). In this case, if the classification images are the same at the two locations, the model processes the d and und locations the same as well, and no ing effect is found. Thus, within

11 36 BRAIN RESEARCH 1080 (2006) Fig. 9 Classification images for PC, a normal observer. See Fig. 10 for details. Fig. 10 Classification images for RP, a normal observer. See Fig. 8 for details.

12 37 Fig. 11 Classification images for SS, a normal observer. See Fig. 8 for details. Fig. 12 Classification images for TB, a normal observer. See Fig. 8 for details.

13 38 BRAIN RESEARCH 1080 (2006) Fig. 13 Classification images for HL, a patient observer. See Fig. 8 for details. Fig. 14 Classification images for CM, a patient observer. See Fig. 8 for details.

14 39 Table 3 Correlations of normal observer and patient classification images with the ideal observer Left location Right location Correlation SE Correlation SE (a) Normal observers EM Cued Signal No signal Und Signal No signal PC Cued Signal No signal Und Signal No signal RP Cued Signal No signal Und Signal No signal SS Cued Signal No signal Und Signal No signal TB Cued Signal No signal Und Signal No signal (b) Patients HL Cued Signal No signal Und Signal No signal CM Cued Signal No signal Und Signal No signal Standard deviations were estimated by Monte Carlo resampling of assumed independent Gaussian-distributed values for the classification images. the constraints of the model, any ing effects found in this simulation reflect differences in the shapes of the classification images at the d and und locations relative to the signal, such that the template at the und location is a poorer match (assuming a positive ing effect). Table 5 and Fig. 18 give the results of the simulations using the ideal templates at the d and und locations and the optimal weighting (w c = 0.80). As there is no difference in the templates, this simulation estimated the ing effect solely based on the differential weighting of the d and und locations and how well the observers' ing effects matched those predictions. Table 6 and Fig. 19 give the results of simulations using the normalized classification images and the optimal weighting (w c = 0.80). In this case, both the weighting and any difference in shape in the classification images may contribute to the ing effect. If the classification images were approximately the same at the d and und locations, then the predicted ing effects for this simulation should be similar predicted ing effects with the ideal templates and the optimal weighting (w c = 0.80). Thus, with the combined results across the simulations, one may assess the relative importance of differential weighting and differential shape at the d and und locations in predicting the observed ing effects. Table 7 summarizes the observed ing effects, the correlations between the d classification image and the und classification on the opposite side, and the P values from the chi-square goodness of fits across the simulations. For the normal observers with ing effects near zero (EM, right s; PC, left s; RP, both s), the 0.50 model had generally good fits, while the other models with the 0.80 weighting (with the observed classification images and the ideal templates) had worse fits that were equally poor. This is an expected result from the lack of a ing effect and also the relatively high correlations of the d and und classification images with the ideal observer. For SS, his classification images had relatively high correlations with each other and the ideal observer, and he showed a ing effect. In his case, the 0.50 model fit poorly, as it should. The 0.80 models had better fits to SS's results for left-sided s and slightly worse fits for right-sided s, and the fits were. These results are consistent with the model of ing effects as differential weighting, with the left- weight close to the optimal 0.80 and the right- weight between 0.50 and A separate simulation with SS's classification images in which the weight was free to vary found that the best-fitting weight was 0.65 (SNR = 1.665, log(crit) = 0.10, χ 2 (0) = 0.004). A suboptimal weighting also describes PC's negative ing effect for right-sided s, as her correlations of the Cued Right and Und Left classification images with the ideal observer and each other were relatively high and the fits to the three models were poor, as expected. A simulation of her classification images with the weights free to vary found that PC's negative ing effect corresponded best to a weight of 0.13 (or 0.87 on the und side) (SNR = 1.530, log(crit) = 0.62, χ 2 (0) = 0.004). EM and TB both had moderate ing effects on the left side, which were fit relatively well assuming 0.80 weights and the ideal templates. Unlike the previous examples, lower correlations to the ideal template were found for the Und Right locations and therefore also for the Cued Left/Und Right correlations. Correspondingly, their ing effects were not fit well with the 0.80 weighting and the observed classification images. Furthermore, the ing effects were fit relatively well assuming equal weighting of the d and und locations and the observed classification images. These results describe ing effects that were determined by a difference in shape only. TB's moderate ing effect for right-sided s were fit well by the 0.80 weighting and the ideal templates. She had a low correlation for the Und Left classification image with the ideal template and with the Cued

15 40 BRAIN RESEARCH 1080 (2006) Fig. 15 Correlations of the ideal observer template (the signal) with the classification images, collapsed over signal presence. Cued L Unc R columns = Cued Left (top left, Figs. 8 14A), Und Right (top right, Figs. 8 14B). Cued R Unc L columns = Cued Right (top right, Figs. 8 14A), Und Left (top left, Figs. 8 14B). Cued = filled squares, und = open circles. Error bars indicate the estimated standard errors from Monte Carlo resampling (see text for details). Fig. 16 Correlations of the d and und classification images collapsed over signal presence. Cued L Unc R columns = Cued Left (top left, Figs. 8 14A) correlated with Und Right (top right, Figs. 8 14B). Cued R Unc L columns = Cued Right (top right, Figs. 8 14A) correlated with Und Left (top left, Figs. 8 14B). Cued = filled squares, und = open circles. Error bars indicate the estimated standard errors from Monte Carlo resampling (see text for details).

16 41 Table 4 Weighted Likelihood model predictions of behavioral performance with classification images as templates, d weight (w c ) = 0.50, by side Cue side Predicted Hv Predicted Hi Predicted FA Predicted ing SNR log(crit) χ 2 (1) P value Normal observers EM Left Right PC Left Right RP Left Right SS Left Right TB Left Right b Patients HL Left Right CM Left Right Performance with left s was predicted by using the Cued Left (top left, Figs. 8 14A) and Und Right (top right, Figs. 8 14B) classification images. Performance with right s was predicted by using the Cued Right (top right, Figs. 8 14A) and Und Left (top left, Figs. 8 14B) classification images. Predicted Hv = predicted valid hit rate. Predicted Hi = predicted invalid hit rate. Predicted FA = predicted false alarm rate. Predicted ing = predicted ing effect (predicted valid hit rate predicted invalid hit rate). SNR = signal-to-noise ratio of stimuli (free parameter). Log(crit) = natural logarithm of the criterion (free parameter). χ 2 (1), P value = chi-square goodness of fit statistic and P value of the predicted and actual hit and false alarm rates, df =1. Right classification image, such that both the 0.50 and the 0.80 weighting with her observed classification images overpredicted her ing effect. The primary result for the patients was a large ing effect with right-sided s. For HL, his right-sided ing effect could be fit with the observed classification images and the equal (0.50) weighting, suggesting that his ing effect could be described by the difference in shape. For CM, note that the negative correlation of the Und Left classification image led to a negative weighting of the und location. His right-sided ing effect was slightly underpredicted with his observed classification images and a weighting of 0.50 and slightly overpredicted with a weighting of Thus, the difference in shape of the classification images was relatively successful in characterizing the general attentional effect shown by neglect patients. With left-sided s, the low correlations for the right und classification images led to poor fits of their modest ing effects (similar to TB's right-sided ing effects). The 0.80 model with the ideal templates predicted HL's ing effect and overpredicted CM's ing effect. Except for TB right-sided ing effects, it appeared that assessing the shape and potential weights of the observed Table 5 Weighted Likelihood model predictions of behavioral performance with ideal observer templates, d weight (w c ) = 0.80, by side Cue side Predicted Hv Predicted Hi Predicted FA Predicted ing SNR log(crit) χ 2 (1) P value Normal observers EM Left Right PC Left Right b RP Left Right SS Left Right TB Left b Right b Patients HL Left Right CM Left b Right See Table 4 caption for details.

17 42 BRAIN RESEARCH 1080 (2006) Table 6 Weighted Likelihood model predictions of behavioral performance with classification images as templates, d weight (w c ) = 0.80, by side Cue side Predicted Hv Predicted Hi Predicted FA Predicted ing SNR log(crit) χ 2 (1) P value Normal observers EM Left Right PC Left Right b RP Left Right SS Left Right TB Left Right Patients HL Left Right CM Left Right See Table 4 caption for details. classification images could describe the ing effects (or lack thereof) of the normal observers. A difference in shape in the classification images also could predict the large right-sided ing effects of the patients. In the cases of modest ing effects and large differences in shape (low correlations of the und classification image with the ideal template and the d classification image), predictions were poor as the models overpredicted the ing Fig. 17 Predicted ing effects from the Weighted Likelihood model (Fig. 3) using the classification images as templates, w c = Cued L Unc R columns = ing effects for left-sided s using the Cued Left (top left, Figs. 8 14A) and Und Right (top right, Figs. 8 14B) classification images. Cued R Unc L columns = ing effects for right-sided s using the Cued Right (top right, Figs. 8 14A) and Und Left (top left, Figs. 8 14B) classification images. Black circles behavioral results (same as Fig. 8). Open circles weight (w c ) = 0.50.

18 43 Fig. 18 Predicted ing effects from the Weighted Likelihood model (Fig. 3) using the ideal templates, w c = Cued L Unc R columns = ing effects for left-sided s using the Cued Left (top left, Figs. 8 14A) and Und Right (top right, Figs. 8 14B) classification images. Cued R Unc L columns = ing effects for right-sided s using the Cued Right (top right, Figs. 8 14A) and Und Left (top left, Figs. 8 14B) classification images. Black circles behavioral results (same as Fig. 8). Gray diamonds weight (w c ) = effect (TB right s, HL and CM left s). Furthermore, in these cases, the ideal weighting (0.80) with the ideal templates could account for the ing effects relatively well. Another aspect of the fits was that the template shape differences could account for ing effects in two normal observers (EM left s, TB left s). In a similar task to the present study with 2-dimensional Gaussian (blurry disk-like in white noise) stimuli, it was found that four observers had no changes in shape in their d and und classification images (Eckstein et al., 2002). As in the cases with the poor fits above (TB right s, HL and CM left s), the ideal weighting and template could account the left-sided ing effects of EM and TB. One possibility is that the correlations of the und classification images were underestimated in these cases (TB and EM). Differential weighting without changing the shape (such as the ideal observer) leads to lower amplitude classification images in the und location (Eckstein et al., 2002). Hence, the correlation of such a lower amplitude classification image with the ideal observer would be more sensitive to noise (both from the observer and the added image noise). Perhaps greater statistical power or higher signal contrasts would lead to better correspondence of the observed classification images with the behavioral performances over all the observers. 3. General discussion and summary Both patients had been assessed initially with a standard battery of tests for hemineglect (Behavioral Inattention Test) and judged to be either partially (HL) or completely (CM) recovered. Despite their significant recoveries, both the behavioral results and the classification images from the present study revealed clear effects consistent with attentional loss in the contralesional field. Thus, it seems that the methodology in this study may be more sensitive at detecting residual attentional loss than the standard testing. There are several differences between the task used in the present study and the battery of tasks used in the Behavioral Inattention Test, which include line bisection, various cancellation tests, and drawing and copying tasks. One difference is that the critical manipulation in the trials in the present study occurred within tenths of seconds, whereas in the standard neglect tests trials tend to be time-unlimited. Thus, it could be that the patients' attentional loss as measured in the present study can only be revealed only under relatively severe time constraints (see also Olk et al., 2002; Harvey et al., 2002). Furthermore, the patients may have been able to use response strategies to compensate for their attentional loss in the standard neuropsychological tests, strategies that could

19 44 BRAIN RESEARCH 1080 (2006) Fig. 19 Predicted ing effects from the Weighted Likelihood model (Fig. 3) using the classification images as templates, w c = Cued L Unc R columns = ing effects for left-sided s using the Cued Left (top left, Figs. 8 14A) and Und Right (top right, Figs. 8 14B) classification images. Cued R Unc L columns = ing effects for right-sided s using the Cued Right (top right, Figs. 8 14A) and Und Left (top left, Figs. 8 14B) classification images. Black circles behavioral results (same as Fig. 10). Gray circles weight (w c ) = not be used in the ing task in the present investigation. Our ability to control for compensatory response strategies might be attributed to the restricted time afforded to the patients when performing our task, as well to the fact that the present task was relatively novel when compared to the standard tests that the patients were possibly more familiar with. Finally, the testing in this study was conducted over 3 to 5 days and therefore represented a relatively intensive and extensive attentional examination when compared to the standard BIT. Thus, the statistical power of our present investigation may have been able to detect and measure the residual attentional deficits that are undetected by more routine tasks that characterize the standardized BIT and everyday life. The behavioral results (ing effects) of the two patients were similar to each other and also consistent with the disengagement deficit pattern that Posner et al. (1984) had found for neglect patients with parietal damage. However, two substantially different patterns of classification images were found in the left hemisphere. When the signal was not present, HL's classification image had no correlation with the signal (ideal observer template). Conversely, when the signal was present, the classification images were positively correlated with the signal. For signal patterns with phase information (such as the checkerboard or Gabor patches), Ahumada (2002) found that nonlinear detection mechanisms that deviate from the assumption that a single perceptual template lead to classification images can lead to this pattern of results: classification images calculated from the signal present trials are biased in favor of the signal, while classification images calculated from the signal absent trials do not exist (i.e., are flat). This is known as location uncertainty, in which the observer is uncertain of the exact location of the signal and can be modeled as the placement of multiple templates at different locations in the stimulus (Pelli, 1985). CM's classification image for the left hemisphere, on the other hand, indicated that he potentially could use the information at that location but appeared to use that information inappropriately under certain conditions. As discussed in Introduction, the ideal observer must have two components, the shape of the template at the location and also the weight given to the information. When either the or the signal appeared on the left, CM's classification image for the left had a high positive correlation with the signal. This suggests that CM could use the information appropriately in this case, both in terms of the shape of the template and the weight given to that information. However, when the appeared on the right side and the signal was not present on the left, the

20 45 Table 7 Summary of model fits of the three models presented in Tables 4 6 and Figs Cue side Cueing effect Correlation d vs. opposite und P values classification images w c = 0.50 P values ideal templates w c = 0.80 P values classification images w c = 0.80 Normal observers EM Left Right PC Left Right b b RP Left Right SS Left Right TB Left b Right b b Patients HL Left Right CM Left b Right Column1: ing effect = valid hit rate invalid hit rate (from Table 2). Column 2: correlations of the dclassification image withthe opposite und classification image (from Fig. 16). Column 3: P values from the chi-square goodness of fits for the model with 0.50 weighting and the observed classification images (from Table 4 and Fig. 17). Column 4: P values from the chi-square goodness of fits for the model with 0.80 weighting and the ideal templates (from Table 5 and Fig. 18). Column 5: P values from the chi-square goodness of fits for the model with 0.50 weighting and the observed classification images (from Table 6 and Fig. 19). correlation with the signal was negative and nearly the same (unsigned) amplitude (see Fig. 5F) as the other classification images. In fact, the classification image suggests that CM took information from the left in favor of signal presence as evidence against signal presence. There are two interpretations to this result. The first is that the high unsigned amplitude indicates that the potential quality of information for the und-no signal classification image was nearly as high as the other classification images, with the reversal in sign suggesting that CM weighted that information entirely inappropriately. Alternatively, it might be said that CM has an entirely inappropriate (opposite) template shape. Thus, for HL and CM, the classification images suggest two different causes of their neglect-type behavior. In general, when the was on the right side of space in this study, CM seemed to have the ability to use (perceive) the information from the left side, but not the ability to give the information the proper importance. Conversely, HL appeared not to able to localize the exact expected signal location on the left side, regardless of location. This is perhaps not an unexpected outcome, as the areas of injury differed, with HL's injury more posterior and dorsal to CM's. While neglect occurs more frequently after right hemispheric brain injury, and particularly to the right parietal cortex, injuries to other areas of the brain also lead to hemineglect. These areas include injury to the frontal cortex (dorsolateral, Critchley, 1966; Heilman et al., 1970, 1983) and injury to subcortical areas, such as the thalamus (Watson and Heilman, 1979; Karussis et al., 2000) and the basal ganglia (Vallar and Perani, 1986; Damasio et al., 1980). Therefore, neglect may be more accurately described as an assembly of syndromes having a similar phenomenology, with possibly different mechanisms of disruption rather than a unitary syndrome (Harvey et al., 2002; Olk et al., 2001). Even within these two patients, substantially different mechanisms of attentional loss appear to be expressed. Perhaps with the testing of more patients, it would be possible to characterize different types of attentional loss (that currently fall under the same category as neglect) to specific areas of brain injury. There is the question of whether the results of the two patients might be due to their age. While the patients did exhibit a general deficit in performance, as indicated by the longer stimulus durations and the higher stimulus contrasts, we do not believe that age accounts for the hemispheric differences in the ing effects and the classification images. First, the patients' ing effects for left-sided s matched well with both the normal observers' ing effects and those predicted by the ideal observer, suggesting that there was not a general discrepancy or loss of the use of the for the patients. Second, it seems unlikely that aging would produce the hemispheric asymmetries consistent with neglect that the patients demonstrated in this study, as there has been no previous suggestion of neglect-like symptoms in normal aged populations (e.g., Lezak, 1995; Gazzaniga, 1998). In conclusion, we tested two patients and five normal observers in a ing task, a common test of visual attention. We found clear indications of attentional deficits for the patients, even after more than a year after their injury and after the patients had partially (HL) or fully (CM) recovered when assessed by a standard neuropsychological examination (the Behavioral Inattention Test). Furthermore, the classification image technique indicated differing causes of neglect for the two patients, even though they showed similar results from behavioral tests (i.e., the ing effects). Like the studies we have conducted with normal observers (Eckstein et al., 2002), the present study suggests that the research technique presented here can test for different types of mechanisms of attentional loss that are otherwise indistinguishable by the standard measures of behavioral performance. Finally, as

21 Fig. 20 A series of horizontal CT image scans for HL showing his lesion. 46 BRAIN RESEARCH 1080 (2006) 26 52

22 47 suggested by these two patients, the types of attentional deficits may differ from patient to patient, and it is hoped that the technique presented here could be used to characterize these different types of deficits with regard to the specific brain areas that are compromised. 4. Experimental procedures We performed a study of classification images in a ing task on 5 normal observers and two parietal patients with a history of neglect. Observers performed a yes/no contrast discrimination of a signal appearing at one of two locations 2.5 to the left and right of a central fixation point (Fig. 1). The signal was a checkerboard pattern configured as a high contrast white X (see Fig. 2, left, for a larger view). As a lower contrast pedestal (contrast = 7.8%) checkerboard appeared in the locations without the signal, this task may be viewed as a discrimination of the lower (pedestal) and higher (signal) contrast checkerboards. The stimulus display appeared for either 50 ms (normals) or 140 ms (patients), with the signal appearing on half the trials, and observers judged on each trial whether the target was present. Immediately prior to the stimulus display, a 2.5 square pre (140 ms) appeared at one of the two possible signal locations; on signal present trials, the pre indicated the signal location with 80% validity. Gaussian image noise was added to the stimuli to calculate classification images; this noise was added to each of the 9 squares comprising the checkerboards (SD = 11.7% contrast). The five normal observers were 4 females, ages 21 (PC, RP, EM) and 22 (TB), and 1 male, age 39 (SS, an author), with no known brain injury or deficit. Each normal observer participated in approximately 2500 trials/observer, and the contrast for their signal was 19.5%, except for TB, who was tested on a slightly higher contrast, 21.9%. The normal observers viewed a CRT monitor in a darkened room with the luminance calibrated by the Dome Calibration system. For the patients, the signal contrasts were higher than those for the normal observers (HL = 27.3%, CM = 23.4%), and the number of trials was less (HL = 820, CM = 1641). The higher contrasts and longer durations (140 ms vs. 50 ms) for the patients were chosen so that overall performances were approximately equivalent to the normal observers. Furthermore, stimuli were presented on a laptop computer with an LCD monitor (Toshiba Satellite 1805-S207), and luminance and color calibrations were performed with the Optical 3.7 (Pantone Colorvision) system Patients Both patients were male (HL, age 69; CM, age 85) and had suffered a right hemispheric stroke at least a year prior to testing. Approximately 1 year and 9 months prior to testing, HL presented with left hemiparesis, left hemianopia, and left neglect and extinction. CT examination (Fig. 20) indicated an extensive area of low density reflecting ischemic damage (without hemorrhage) in the territory of the right posterior cerebral artery, especially the striate and peristriate cortices, posterior temporal neocortex, and posterior periventricular white matter, with some extension into the inferoposterior parietal area. HL also suffered a lacunar infarct in the right caudate, as well as some deep ischemic changes to white matter. On the Behavioral Inattention Test (BIT, a battery of cancellation, bisection, drawing, and copying tasks designed to assess neglect, with any score below 129 out of 146 indicating neglect, Wilson et al., 1987), HL scored 100/146 approximately 3 months after his injury, indicating a clear deficit. Just prior to testing, his score on the BIT had improved to 126 but still indicated residual neglect at the time of testing. Approximately 1 year and 4 months prior to testing, CM presented with a left hemiparesis, left neglect, and left hemisensory deficit, including the left arm, hand, wrist, and leg. His MRI image scan immediately after the stroke indicated an ischemic infarction of the right middle cerebral artery (MCA); also, an old right MCA territory infarction was found, apparently from a transient incident 2 years prior to the acute stroke and from which he apparently had made a full functional recovery (see Fig. 21). At the time of testing, CM's score on the BIT was 141/146, suggesting a nearly full recovery from neglect (as assessed by the BIT) Calculation of classification images Classification images were calculated for each side (left and right) based on presence on that side and were further divided by signal presence on that side. Thus, for each side, there were six classification images: d, d-signal, dno signal, und, und-signal, and und-no signal. The individual noise fields were weighted by the observer's response on that trial, either 1 for a no response or +1 for a yes response. This weighting is modified from a simple rule for combining noise fields across different response outcomes when calculating classification images (Beard and Ahumada, 1998; Ahumada, 2002; Murray et al., 2002). The classification image for each of the six outcomes is the mean of the individual noise fields (after weighting) corresponding to that outcome. One-sample Hotelling T 2 statistics were employed to assess whether the classification images were nonzero (different from a flat or zero value classification images) or differed from the ideal observer; two-sample Hotelling T 2 statistics were employed to assess differences between classification images. The Hotelling T 2 statistic is the multivariate equivalent to the t statistic (see Appendix A). Correlations of the observed classification images with the signal were calculated to assess the match of the observed templates to the ideal template in this task. As a method to assess the relationship between the classification images and the observed ing effects, correlations were also calculated between the classification images when one side was d and the classification images when the opposite side was not d. Standard errors of the correlations were estimated by a simulation method known as Monte Carlo resampling (Efron and Tibshirani, 1993). The values of each square of the classification images were assumed to be independent and Gaussian-distributed with the same means and standard deviations as the empirical values.

23 48 BR A IN RE S EA RCH ( )