Participants were recruited at the Department of Radiology at University of California, Los Angeles (UCLA) Medical Center from 3 volunteer subject populations: practicing thoracic radiologists (n
7; 3 females; mean age = 51.6), fourth-year radiology residents (n
7; 3 females; mean age = 30.9), and first-year radiology residents (n
7; 2 females; mean age = 28.6). Practicing radiologists had at least 10 years experience postresidency (mean = 18.9 years). One fourth-year resident had large uncorrectable motion artifacts in his fMRI data and was excluded from all further analyses. In general, attrition rates are low across residency in Radiology. All participants had normal or corrected-to-normal vision. Informed consent was received from each participant, and all experimental procedures were approved by the UCLA Office for the Protection of Research Subjects.
Data were collected inside a magnetic resonance imaging (MRI) device while participants underwent functional scans of their brains. Stimuli were displayed on MR-compatible goggles controlled by a Macintosh frame buffer. The goggle display subtended 31 degrees of visual angle horizontally and 23 degrees vertically. While viewing stimuli, participants indicated their responses with an MR-compatible button box equipped with 4 buttons.
Stimuli were normal and abnormal chest radiographs obtained from a commercially available CD published by the Japanese Society of Radiological Technology (JSRT) in cooperation with the Japanese Radiological Society. Each abnormal radiograph contained a single lung nodule, a potentially malignant round lesion located in the lung field. The abnormal radiographs were further divided by the JSRT publishers into 3 difficulty levels. An expert thoracic radiologist independently evaluated each image and excluded images with the most difficult-to-detect nodules and those of poor image quality. This resulted in a usable set of 92 normal and 100 abnormal radiographs.
Two versions of each radiographic image were used in the main part of the experiment: intact and scrambled (). The 2 radiograph types, normal and abnormal, were crossed with the 2 image types, intact and scrambled, to yield 4 total stimulus conditions. Scrambled radiographs were created by dividing each image into 25 squares and randomly shuffling the location of all but one of those squares. The one square that remained in its original location in each image was, for the abnormal radiographs, the square that contained the lung nodule and, for the normal radiographs, the square that was cued as a possible nodule location. Cue locations for the normal radiographs were matched to actual nodule locations in the abnormal radiographs. To ensure that each cue was a plausible nodule location, cue location assignment for the normal radiographs was determined by a postresidency radiologist at UCLA Medical Center who did not participate as a subject in the study.
Because scrambling the radiographs introduced vertical and horizontal lines in the images, grids were drawn on all stimuli, both intact and scrambled. Each lung nodule was entirely contained within a square and never obscured by the grid. For sample intact and scrambled nodule-containing radiographs, see . All images were square and subtended a visual angle of approximately 19 degrees.
Each participant completed one 80-min scanning session followed by an expertise posttest completed outside of the scanner. The scanning session contained several anatomical scans, 3 localizer scans, and 3 rapid event-related scans. Localizer scans were used to define regions of interest. Examples of images used in the localizer scans are shown in .
To define visual cortical areas selective for processing faces, participants viewed 8 blocks of faces in alternation with 8 blocks of objects. Note that rest or fixation blocks were not placed between the blocks of faces and objects. Nine images were shown per stimulus block; 72 faces and 72 objects were displayed in total. Each image was presented for 1.7 s with 0.3-s interstimulus intervals yielding a block duration of 18 s. To control for attention, participants performed a 1-back task. On one in every 9 stimulus presentations, on average, the same image was shown on successive trials. Participants were instructed to press a button when an image repeated in this fashion. Performance on the one back task was 81% correct for novices and 85% correct for experts; these means did not differ reliably.
To identify regions of object selective cortex, including LO, that might be specialized for processing radiographs, participants viewed 8 blocks of intact radiographs in alternation with 8 blocks of scrambled radiographs. Note again, that rest or fixation blocks were not placed between the blocks of intact and scrambled radiographs. Scrambled radiographs for the localizer scan were finely scrambled; each image contained 100 equal-size squares (). Nine images were shown per stimulus block; 32 unique radiographs were shown during the localizer scan, repeated as necessary to yield a total of 72 intact presentations and 72 scrambled presentations. Each image was presented for 1.7 s with 0.3-s interstimulus intervals yielding a block duration of 18 s. Radiographs shown in the localizer scan were not shown in any event-related diagnosis scans. To control for attention, participants performed a 1-back task as described above for the face localizer scan (note that participants did not perform a diagnosis task on these images). Performance on the one back task was 96% correct for novices and 99% correct for experts; these means did not differ reliably.
We also employed standard retinotopic mapping procedures (Engel et al. 1994
; Sereno et al. 1995
; DeYoe et al. 1996
) to identify retinotopic visual areas (e.g., V1, V2, and V3). Participants fixated on a square positioned in the center of the display. A wedge filled with a high contrast temporally reversing checkerboard pattern rotated around fixation, completing one rotation every 30 s. To keep participants fixated and attending, the fixation square intermittently changed from black to white or vice versa. Participants were instructed to press a button as rapidly as possible each time fixation changed color.
The diagnosis scans were event-related scans in which participants viewed intact and coarsely scrambled radiographs and judged whether a cued region in each radiograph contained a lung nodule (). A new trial occurred every 3.0 s. The sequence of a single trial was as follows: 1) a grid the same size as the radiographs was displayed for 1000 ms. The grid contained a fixation cross in the unit of the grid that the subject was to judge. Participants were instructed to move their eyes to the fixation cross and then to remain fixated at that location for the remainder of the trial. 2) The grid, but not the fixation cross, remained on the screen and a radiograph was displayed for 500 ms. 3) The radiograph was removed, but the grid and fixation remained for 1000 ms.
Each scan consisted of 124 trials composed of 25 radiographs in each of the 4 stimulus conditions: 1) intact, normal; 2) scrambled, normal; 3) intact, abnormal; 4) scrambled, abnormal, and 24 rest trials in which no radiograph was presented. Condition order was counterbalanced using an m-sequence (Buracas and Boynton 2002
Functional MRI data were acquired using a blood oxygen level–dependent contrast-weighted echo-planar pulse sequence (3T; time echo
25; time repetition
3 s; flip angle
90; field of view
20 cm; voxel size
4 mm; 36 slices parallel to the anterior-posterior commissure line). High-resolution conventional anatomical images were acquired coplanar to the functional data, and T1
-weighted volumetric scans were acquired for cortical flattening.
fMRI Data Analysis
Localizer scans were analyzed by simple correlation of each voxel's time series with a sinusoid at the stimulus frequency. Because the face localizer and radiograph localizer scans did not include rest blocks, this correlation effectively measured the reliability of the difference between activity in the 2 conditions (faces vs. objects or intact vs. scrambled localizers). Active pixels were identified as those with a correlation above a threshold whose activity was in-phase with the stimulus. Different threshold levels were tested for area identification; the overall pattern of results did not depend upon threshold.
Event-related scans were analyzed using the general linear model and the “Finite Impulse Response” approach. The design matrix contained regressors for each time point in the response for each condition. Estimates for each regressor were computed using ordinary least squares from the average timecourse of active voxels in each visual area. The amplitude of the response for each condition was computed as the peak of the estimated impulse response, excluding the first and last time points. This peak was computed independently for each subject, and so the time point at which it occurred varied slightly between subjects. Other estimates of response amplitude (e.g., area under the curve and fit model hemodynamic response) yielded similar results.
Following standard techniques, the FFA was identified as voxels in the mid-fusiform gyrus that were active in the face localizer scan (Kanwisher et al. 1997
). The occipital face area (OFA) was identified as voxels on the posterior lateral surface of the occipital lobe that were active in the face localizer scan. Areas on the lateral aspect of the occipital lobe that were consistently active in the radiograph localizer scan were defined as LO. Additional contiguous areas on the ventral surface of the occipital lobe were also identified; this region corresponds to the posterior fusiform gyrus and has been annotated pFus as in other studies. Note that the regions in LO and pFus identified in this way correspond to only a subset of the larger lateral occipital complex that is often identified by comparing images of other types of objects to textures (Malach et al. 1995
Correlating Behavior and Neural Activity
We computed Pearson correlation coefficients between participants’ estimated response amplitudes and behavioral performance on the diagnosis task. Corresponding Student's t values were also computed. To test whether our results were due to low-performing subjects, we removed subjects from the analysis when their performance was below a d′ of 0.75. Two subjects were below this threshold for scrambled radiographs only; these data were removed from the correlation analysis and only data for intact radiographs (both activity and performance) were used. One subject was below d′ of 0.75 for both intact and scrambled objects, and their data were completely removed from the analysis.
To compare correlations between groups, we conducted a randomization analysis in which participants were assigned to groups at random, and the correlation between behavior and FFA activity was computed for each group. The difference between group correlations was then computed. This randomization was repeated 1000 times, yielding a null distribution of group correlation differences from which p values were computed.
General Expertise Posttest
To provide a measure of general radiology expertise, each participant completed a posttest outside of the scanner. The posttest consisted of 15 chest radiographs selected from radiology board certification training images. Test items ranged in degree of diagnosis difficulty, for example, pneumopericardium, aortic aneurysm, and mitral valve calcification. Participants were allowed to view each image freely under no time constraints and were asked to provide a written diagnosis for each. An expert radiologist at UCLA Medical Center who was not a participant in the study scored expertise posttests.