AJR 2003; 180:607-619
© American Roentgen Ray Society
Fundamentals of Clinical Research for Radiologists |
Visualizing Radiologic Data
Stephen J. Karlik1
1 Department of Diagnostic Radiology and Nuclear Medicine, Rm. 2MR21, University
of Western Ontario, London Health Sciences Center-University Campus, 339
Windermere Rd., London, Ontario N6A 5A5, Canada.
Received June 13, 2002;
accepted after revision July 9, 2002.
Address correspondence to S. J. Karlik.
Series editors: Craig A. Beam, C. Craig Blackmore, Steven Karlik, and
Caroline Reinhold.
This is the ninth in the series designed by the American College of
Radiology (ACR), the Canadian Association of Radiologists, and the
American Journal of Roentgenology. The series, which will ultimately
comprise 22 articles, is designed to progressively educate radiologists in the
methodologies of rigorous clinical research, from the most basic principles to
a level of considerable sophistication. The articles are intended to
complement interactive software that permits the user to work with what he or
she has learned, which is available on the ACR Web site
(www.acr.org).
Project coordinator: Bruce J. Hillman, Chair, ACR Commission on Research
and Technology Assessment.
Introduction
It should come as no surprise to radiologists, who earn their living by
analysis of visual information, that the analysis and presentation of
scientific data should also have a significant visual component. Not only does
the visual presentation enhance the clarity of the data, whether for
presentation or for publication, but in fundamental ways it also assists in
our understanding of it. In fact, modern data graphics should be considered
instruments for reasoning about quantitative information. Sometimes the best
way to understand, describe, or summarize numeric data is to look at a picture
of it. In consideration of statistical graphics, one publication rises above
all others: The Visual Display of Quantitative Information, by E. R.
Tufte [1]. Far from being a
cold and clinical tome, as the title suggests, this is an entertaining and
approachable work from which we can take important concepts and apply them to
the expression of radiologic data.
The display of data in graphs, charts, and diagrams has a specific aim: to
discover any patterns in the data. Although this data display is particularly
useful in continuous data (described in "Exploring and Summarizing
Radiologic Data" in this series
[2]), this article will use
examples to illustrate those instances in which other graphic styles can help
us conceptualize the phenomena underlying our observations. When data are
prepared for publication, it is customary to choose the most relevant and
smallest number of illustrations or radiographic images to describe the
findings. Statistical graphs can assist in this process by revealing and
concentrating large amounts of data into a manageable size for portrayal.
Graphic excellence has certain properties: clarity, precision, efficiency,
consideration of several variables simultaneously, and honesty in revealing
the data [1].
Graphic Integrity
This article is an examination of a series of figures from the recent
radiology literature, with careful attention to the basis of graphic integrity
as outlined by Tufte [1].
Graphic integrity includes using the physical size of numbers or symbols in
proportion to the actual values; showing data variation—not design
variation; using clear and unambiguous labeling; not quoting data out of
context; and avoiding having the number of graphic dimensions exceed the
number of dimensions in the data
[1]. Other key definitions and
concepts brought up by Tufte are illustrated in the figures. Because these
figures are reprinted to illustrate points about graphic design, none was
changed to conform to the American Journal of Roentgenology style for
figures.
One fundamental concept in judging graphic competence is that of
"data ink," in which the data-ink ratio equals the ink used for
data (data ink) divided by the total ink used in the graph
[1]. Therefore, background
grids, three-dimensional pictures, shading, and hyperactive bar fills are
unproductive ink, diluting the data-ink ratio. For clarity, then, nondata ink
should be erased. The overall principles to optimize the data-ink ratio
include showing the data, maximizing the data-ink ratio, erasing nondata ink,
and erasing redundant data ink
[1]. Some of the radiologic
examples in this article pertain to the issue of data ink.
Furthermore, there are annoying charts and graphs that substitute graphic
variation for data variation. One type of colorfully named
"chartjunk" [1] is
the moiré optical effect caused by closely spaced lines. You have seen
this effect on the television screen (particularly in striped clothing), and
now, with the promulgation of computerized graphing programs, it is becoming
more common in research reports. Although background grids can assist in the
reading of a complex data set, de-enhancing the grid to a lighter shade of
gray may help to minimize the optical assault. Most of the data ink should be
devoted to data variation. Following this premise enhances the efficiency of
communication. In the design of statistical graphs, the ability to portray
complexity, structure, and density of data should always be considered.
The Good, the Bad, and the Ugly
This article reviews 21 figures taken from the recent radiologic literature
to examine how these graphic presentation issues have been dealt with. Figures
1A,
1B and
2 are examples of data-intense
multivariable graphs in which a substantial amount of data is concentrated in
a format that permits visualization of data variation in patients (Fig.
1A,
1B) and temporal relationships
(Fig. 2). Figure
1A,
1B has a high data-ink ratio
and allows the reader to easily comprehend the control-versus-patient
differences in the MR imaging determination of parotid gland size.
Figure 2, although containing a
background grid that dilutes the data-ink ratio somewhat, is effective in
coordinating the temporal events associated with contrast enhancement.
A previous article in this series discussed graphing two variables to show
a relationship in their correlation
[2]; the graph style shown in
Figure 3 allows the comparison
of two characteristics. Two phenomena that are different but related are
plotted with displacement on the left y-axis and velocity on the
right y-axis. The graphic is data intensive with a high data-ink
ratio. Unfortunately, the choice of the x-axis position has obscured
some of the x-axis tick labels and modestly confuses the
interpretation of the data.
View larger version (27K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 3. —Example of figure that successfully illustrates temporal
relationship between two dependent variables. Graph shows plotting
relationship between two different but related phenomena using two different
y axes: displacement (on left) and velocity (on right) for mean
through-plane motion of prosthetic valve. This figure has high data-ink ratio,
especially with error bars included. Choice for x-axis position is
compromised, leading to some obscuring of data values and of x-axis
tick labels. (Reprinted with permission from
[5])
|
|
Figure 4 shows
box-and-whiskers plots for contrast-to-noise ratios for a variety of different
coronary vessel segments. No statistical differences exist, and the plot
permits the reader to visualize that result. However, the plot does not give
the number observed for each artery segment and contains additional lines of
division that are nondata ink and could be erased.
View larger version (24K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 4. —Box-and-whiskers plots. Graph shows contrast-to-noise ratio
in electron-beam CT coronary angiography for different coronary vessel
segments. Bottom and top edges of box are 25th and 75th percentiles,
horizontal line represents the median, and error bars delimit extent of 10th
and 90th percentiles. No statistical differences were observed, and this type
of plot effectively portrays this data variability. LM = left main coronary
artery, LAD = left anterior descending coronary artery, LCX = left circumflex
coronary artery, RCA = right coronary artery, p = proximal segment, m = middle
segment, d = distal segment. (Reprinted from
[6])
|
|
Figure 5 shows an example of
the ubiquitous receiver operating characteristic curve. In this instance,
however, the straightforward curve with 10 data points is obscured in a sea of
nondata ink, including the background grid, line of unity, extra axis tick
marks, and the inserted legend, which is clearly not needed because only one
data set is plotted on the graph. All these represent chartjunk and should be
eliminated. Compare Figure 5
with Figure 6A,
6B, in which four curves are
plotted and the data-ink ratio is high. It is clear from the curves that no
differences were seen for the four display formats and three abnormalities
(a-c). These data could be presented in tables because no significant
differences were observed.
View larger version (30K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 5. —Example of poor data-ink ratio for receiver operating
characteristic curve. Graph shows only 10 data points, which are obscured by
tremendous amount of nondata ink, including background grid, tick marks, and
line of unity. DAFL = differential air-fluid level. (Reprinted from
[7])
|
|
View larger version (23K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 6A. —Example of data that could have been handled in table format.
(Reprinted with permission from
[8]) Graphs show findings for
reticular (A), small nodular (B), and ground-glass (C)
abnormalities in four display formats. Appropriate receiver operating
characteristic curves are used, but curves are not significantly different for
any abnormalities. Repetition is unproductive. In each graph, it is difficult
to discern individual curves and their identification.
|
|
View larger version (24K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 6B. —Example of data that could have been handled in table format.
(Reprinted with permission from
[8]) Graphs show findings for
reticular (A), small nodular (B), and ground-glass (C)
abnormalities in four display formats. Appropriate receiver operating
characteristic curves are used, but curves are not significantly different for
any abnormalities. Repetition is unproductive. In each graph, it is difficult
to discern individual curves and their identification.
|
|
View larger version (22K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 6C. —Example of data that could have been handled in table format.
(Reprinted with permission from
[8]) Graphs show findings for
reticular (A), small nodular (B), and ground-glass (C)
abnormalities in four display formats. Appropriate receiver operating
characteristic curves are used, but curves are not significantly different for
any abnormalities. Repetition is unproductive. In each graph, it is difficult
to discern individual curves and their identification.
|
|
Figure 7 is our first
example of moiré optical vibrations. This complex figure describes the
calculated optimum treatment strategy in a two-way sensitivity analysis
varying the relative risk of failure after stent placement. The graph shows a
decrease in relative risk with the enlarging proportion of patients requiring
stent placement after percutaneous transluminal angioplasty. However, no
confidence intervals are shown, and the input confidence interval and
proportion indicated by the arrows suggest that the three groups may not be
differentiated.
View larger version (21K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 7. —Figure in which data-ink ratio and optical vibrations
(moiré effect) are poor. Graph shows complex theoretic analysis of
optimal treatment strategy using two-way sensitivity analysis. PTA =
percutaneous transluminal angioplasty, SS = selective stent placement, CI =
confidence interval. (Reprinted with permission from
[9])
|
|
Moiré effects (optical noise) are one design fault in
Figure 8. Additional chartjunk
is seen in the threefold repetition of the type of radiologists and the actual
data values sitting atop each bar. An examination of the amount of data
actually shown in the figure reveals very few data points considering the
amount of ink used to represent them. Also, no significant differences exist
in detectability between any measurements for any of the lesion types. A
re-plot of the data values that decreases the repetition and clearly portrays
the paucity of data (Fig. 9)
still shows a lack of significance.
View larger version (45K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 8. —Example of figure that could have been simplified. Bar chart
shows average detectability of lung abnormalities divided into severity for
two groups of radiologists and four display methods. The presentation has two
principal problems: moiré vibrations (optical noise) and redundancy,
with the two groups of radiologists repeated for each degree of abnormality.
Reprinted with permission from
[8])
|
|
The bar charts shown in Figure
10 are also dominated by optical effects. The graphs depict the
area under the receiver operating characteristic curve for 20 radiologists
interpreting from four different displays. This figure occupies a considerable
amount of visual real estate to show virtually no significant differences.
Although minimal differences are indicated, no correction for multiple
comparisons is indicated nor are confidence intervals shown.
View larger version (66K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 10. —Example of bar charts dominated by moiré patterns.
Illustration of all raw data for many areas from receiver operating
characteristic analyses hides fact that multiple comparisons would require
additional statistical tests. There is little value in occupying so much
visual real estate for not much significant data. Reprinted with permission
from [8])
|
|
The three panels of Figure
11A,
11B,
11C show mild moiré
patterns and background grids. The graphs illustrate the decrease in number of
vertebral disks seen with a decrease in radiation dose. However, no
statistical tests were indicated. Normally, it is sufficient to plot only the
upgoing section of the error bars on the top of each bar. However, in this
instance, the error bars are actually the data range (the same as the range
whiskers in the box-and-whiskers plot), so this is an unfamiliar hybrid
plot.
View larger version (37K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 11A. —Examples of moiré patterns associated with bars filled
with opposing hash lines (each representing a different observer) and effect
of including nondata ink (grids). Values for p are not indicated.
(Reprinted with permission from
[10]) Bar charts show findings
in lung-equivalent (A), heart-equivalent (B), and
sub-diaphragm-equivalent (C) regions.
|
|
View larger version (33K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 11B. —Examples of moiré patterns associated with bars filled
with opposing hash lines (each representing a different observer) and effect
of including nondata ink (grids). Values for p are not indicated.
(Reprinted with permission from
[10]) Bar charts show findings
in lung-equivalent (A), heart-equivalent (B), and
sub-diaphragm-equivalent (C) regions.
|
|
View larger version (31K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 11C. —Examples of moiré patterns associated with bars filled
with opposing hash lines (each representing a different observer) and effect
of including nondata ink (grids). Values for p are not indicated.
(Reprinted with permission from
[10]) Bar charts show findings
in lung-equivalent (A), heart-equivalent (B), and
sub-diaphragm-equivalent (C) regions.
|
|
Figure 12A,
12B shows the upward shift in
receiver operating characteristic area values observed when a group of
radiologists used computer-aided diagnosis. The choice of bar fills does not
dominate the visual picture. The shift to higher receiver operating
characteristic areas is clearly seen, and the variability of the distribution
in performance is intact and interpretable. The use of filled versus open bars
is an effective method of delineation between groups in
Figure 13. However, the graph
does contain superfluous background grids, and design variation was chosen
over data variation. One of the rules for graphic design suggested by Tufte
[1] is that the graph's
dimensions should not exceed the data dimension. Here we have a
three-dimensional plot of only two-dimensional data. The graph design adds
substantial visual ink without adding anything to the interpretation. However,
unless the graphs are carefully considered, even one with copious data ink can
be confusing.
View larger version (81K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 12A. —Example of figure that provides value-filled expression of
improvement in diagnostic accuracy and leaves variability visible. (Reprinted
with permission from [11]) Bar
charts show diagnostic accuracy without (A) and with (B)
computer-aided diagnosis (CAD). Bars have muted moiré effect and charts
have more pleasing overall appearance compared with those of Figures
8,
10, and
11A,
11B,
11C. Panel B shows that
using CAD resulted in increase in diagnostic accuracy for all groups of
radiologists.
|
|
View larger version (65K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 12B. —Example of figure that provides value-filled expression of
improvement in diagnostic accuracy and leaves variability visible. (Reprinted
with permission from [11]) Bar
charts show diagnostic accuracy without (A) and with (B)
computer-aided diagnosis (CAD). Bars have muted moiré effect and charts
have more pleasing overall appearance compared with those of Figures
8,
10, and
11A,
11B,
11C. Panel B shows that
using CAD resulted in increase in diagnostic accuracy for all groups of
radiologists.
|
|
View larger version (101K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 13. —Example of visually effective use of filled versus open bars
for comparing distribution of number of cases per channels visualized. Use of
three-dimensional bars gives graphic variation but adds no value to depiction
of data. Figure also has nondata ink in background. (Reprinted with permission
from [12])
|
|
Figure 14 shows the raw and
mean values for a number of measures of FDG uptake for 10 patients. The reader
cannot follow the actual values from each patient because too many overlapping
symbols appear. Although the mean (the only filled symbol) is easy to pick
out, the error bars add to the confusion. Clutter could have been avoided by
offsetting the mean and standard deviation plots to the side of the raw
data.
View larger version (17K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 14. —Example of complicated scatterplot. Figure depicts large
amount of information for variety of FDG parameters for 10 patients. It is
difficult to follow specific values for individual patients and to discern
mean percentage differences (•). Error bars are confusing. (Reprinted
with permission from [13])
|
|
Figure 15A,
15B shows the alterations in
proportion for a group of 20 radiologists interpreting images from two
formats. They were given three types of images to view and asked which gave
the best processing. No significant differences were reported. Although the
bar graphs show interradiologist variability well, much ink is used to show an
absence of significant changes between formats. Because all the bars add up to
unity (one), the black infill for the third proportion is redundant.
View larger version (61K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 15A. —Example of interesting use of data ink to show proportions
for two variables and 20 observers, with change in display parameter. No
difference exists in discrimination between modalities; therefore, much ink is
used to show no differences. (Reprinted with permission from
[8]) Bar charts show
differences in observer interpretation of nonzooming (A) and twofold
zooming (B) soft-copy displays.
|
|
View larger version (62K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 15B. —Example of interesting use of data ink to show proportions
for two variables and 20 observers, with change in display parameter. No
difference exists in discrimination between modalities; therefore, much ink is
used to show no differences. (Reprinted with permission from
[8]) Bar charts show
differences in observer interpretation of nonzooming (A) and twofold
zooming (B) soft-copy displays.
|
|
Kaplan-Meier curves (Fig.
16A,
16B) are rarely seen in
radiology but are common in clinical studies. These curves are excellent for
showing how rapidly a proportion of different populations reaches a
predetermined clinical outcome (in this instance, stroke) for two populations
divided by sonographic criteria on day 0. The left panel represents less than
50% stenosis and the right panel, greater than 50% stenosis. It is unfortunate
that the two panels have different y-axis ranges. Visually, it
appears that the patients with nonhypoechoic findings in the greater-than-50%
group have about the same number of strokes as both groups in the
less-than-50% panel on the left. They appear about equal, however, because of
the change in scale between the panels.
View larger version (19K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 16A. —Example of Kaplan-Meier survival graphs. (Reprinted with
permission from [14]) Graphs
illustrate proportion of individuals who remain without stroke divided by
degree of stenosis of less than 50% (A) and greater than 50%
(B). Each group is further divided by nonhypoechoic and hypoechoic
findings. Although patients with nonhypoechoic findings in B have
higher occurrence of strokes than those of both groups in A, difference
in y-axis range in B makes proportions appear nearly
identical.
|
|
View larger version (21K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 16B. —Example of Kaplan-Meier survival graphs. (Reprinted with
permission from [14]) Graphs
illustrate proportion of individuals who remain without stroke divided by
degree of stenosis of less than 50% (A) and greater than 50%
(B). Each group is further divided by nonhypoechoic and hypoechoic
findings. Although patients with nonhypoechoic findings in B have
higher occurrence of strokes than those of both groups in A, difference
in y-axis range in B makes proportions appear nearly
identical.
|
|
Changing the axis to accentuate the differences between groups is also
shown in Figure 17A,
17B,
17C. Whereas the left panel
has time points for 30, 60, and 90 sec, the center and right panels show the
same data for one time point only, and the three scatterplots have different
axis ranges. The inclusion of all the raw data is commendable, but no
indication of statistical differences is shown. Using one graph could have
eliminated repetition, and additional lines could have joined the same tumor
at each time point to show whatever trends were found in the temporal
evolution of the enhancement.
View larger version (18K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 17A. —Three scatterplots showing attenuation of early-enhanced CT
images of adenomas and nonadenomas at different times after injection of
contrast material. No statistical differences were indicated. (Reprinted with
permission from [15])
Scatterplots show data at different time intervals: 30, 60, and 90 sec
(A); 180 sec only (B); and 30 min only (C). Because
y-axis scales are changed for each part, this presentation visually
suggests that discrimination between groups is noted at 30 min. Parts B
and C should have also been plotted with attenuation versus all times
of observation to reduce redundancy and nondata ink.
|
|
View larger version (15K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 17B. —Three scatterplots showing attenuation of early-enhanced CT
images of adenomas and nonadenomas at different times after injection of
contrast material. No statistical differences were indicated. (Reprinted with
permission from [15])
Scatterplots show data at different time intervals: 30, 60, and 90 sec
(A); 180 sec only (B); and 30 min only (C). Because
y-axis scales are changed for each part, this presentation visually
suggests that discrimination between groups is noted at 30 min. Parts B
and C should have also been plotted with attenuation versus all times
of observation to reduce redundancy and nondata ink.
|
|
View larger version (13K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 17C. —Three scatterplots showing attenuation of early-enhanced CT
images of adenomas and nonadenomas at different times after injection of
contrast material. No statistical differences were indicated. (Reprinted with
permission from [15])
Scatterplots show data at different time intervals: 30, 60, and 90 sec
(A); 180 sec only (B); and 30 min only (C). Because
y-axis scales are changed for each part, this presentation visually
suggests that discrimination between groups is noted at 30 min. Parts B
and C should have also been plotted with attenuation versus all times
of observation to reduce redundancy and nondata ink.
|
|
Figure 18A,
18B,
18C shows three-dimensional
graphs for four spectroscopic measurements and clinical outcome in three
groups of patients. Three-dimensional graphs are intuitively difficult to
comprehend and these examples also show moiré effects. The use of three
graph dimensions is appropriate to the three data dimensions: proportion,
MR-spectroscopy measurement, and clinical outcome. The number of patients
whose findings contribute to each of the bars is small, however, so this graph
format over-states the value of the data. Also the lack of confidence
intervals allows the graph to appear to tell a definitive story, whereas the
variability of the data that would be associated with such low numbers is not
illustrated. Similarly, the three-dimensional bar graphs for MR imaging
findings in Figure 19A,
19B,
19C show an appropriate number
of dimensions (three: grade, cohort, and age). No statistical analysis is
indicated nor are confidence intervals shown. It appears that the three grades
of the three panels increase with age in the whole cohort independent of the
group subdivision. A considerable amount of visual real estate is used to
illustrate data that have a common pattern. The findings from these three
graphs could be summarized in a few sentences in the results section of the
text.
View larger version (49K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 18A. —Example of complicated three-dimensional bar graphs that are
difficult to understand. Moiré effects are present also. (Reprinted
with permission from [16])
Graphs illustrate complex relationships between four measures and clinical
out-come for three groups of patients: neonates (A), children
(B) infants (C). Graphs appear to hold substantial amount of
information, but close examination reveals that each bar represents few
individuals and findings are visually overstated. This combination of
moiré effects and complex data presentation makes data difficult to
apprehend.
|
|
View larger version (92K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 18B. —Example of complicated three-dimensional bar graphs that are
difficult to understand. Moiré effects are present also. (Reprinted
with permission from [16])
Graphs illustrate complex relationships between four measures and clinical
out-come for three groups of patients: neonates (A), children
(B) infants (C). Graphs appear to hold substantial amount of
information, but close examination reveals that each bar represents few
individuals and findings are visually overstated. This combination of
moiré effects and complex data presentation makes data difficult to
apprehend.
|
|
View larger version (66K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 18C. —Example of complicated three-dimensional bar graphs that are
difficult to understand. Moiré effects are present also. (Reprinted
with permission from [16])
Graphs illustrate complex relationships between four measures and clinical
out-come for three groups of patients: neonates (A), children
(B) infants (C). Graphs appear to hold substantial amount of
information, but close examination reveals that each bar represents few
individuals and findings are visually overstated. This combination of
moiré effects and complex data presentation makes data difficult to
apprehend.
|
|
View larger version (60K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 19A. —Example of material that could have been presented in text or
table format because no significant differences were found and data content is
minimal. (Reprinted with permission from
[17]) Three-dimensional graphs
show grade-scoring changes for subgroups sulcai (A), ventricular
(B), and white matter (C) grades and ages. No error bars are
shown, and numbers of subjects in each subgroup are not given. CHS =
cardiovascular health study, NF = nonblack female, BF = black female, NM =
nonblack male, BM = black male.
|
|
View larger version (56K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 19B. —Example of material that could have been presented in text or
table format because no significant differences were found and data content is
minimal. (Reprinted with permission from
[17]) Three-dimensional graphs
show grade-scoring changes for subgroups sulcai (A), ventricular
(B), and white matter (C) grades and ages. No error bars are
shown, and numbers of subjects in each subgroup are not given. CHS =
cardiovascular health study, NF = nonblack female, BF = black female, NM =
nonblack male, BM = black male.
|
|
View larger version (55K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 19C. —Example of material that could have been presented in text or
table format because no significant differences were found and data content is
minimal. (Reprinted with permission from
[17]) Three-dimensional graphs
show grade-scoring changes for subgroups sulcai (A), ventricular
(B), and white matter (C) grades and ages. No error bars are
shown, and numbers of subjects in each subgroup are not given. CHS =
cardiovascular health study, NF = nonblack female, BF = black female, NM =
nonblack male, BM = black male.
|
|
An odd combination of two measurements in one graph is seen in
Figure 20. The main panel
represents the mean and 95% confidence intervals for the loss of cartilage
thickness under pressure for 210 min. The inserted panel has a different time
axis, although the scale is the same. Perhaps a better way to show these data
would be to use the release point at 210 min as the zero point with times
negative before (during compression) and times positive afterward (during
decompression). The two y-axes should be either the same or better
coordinated.
View larger version (19K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 20. —Unusual figure that inserts graph of completely different
phenomenon within main (enclosing) graph. Although it is sometimes useful to
have different plots using different axes in one figure, this combination is
both confusing and potentially misleading. Minimum acceptable figure would
have identical time axis, perhaps with release point at which time equals
zero. (Reprinted with permission from
[18])
|
|
The final figure of this review, Figure
21A,
21B, has two panels showing
the change in two phenomena as a function of time after angioplasty. The
graphs have a good data-ink ratio, and actual measurements for 10 patients are
illustrated. Although the overall patterns can be discerned, the mean values
(dashed lines) are partially obscured, and the line indicating abnormal values
is also a dashed line. The y-axis on panel b has been broken between
6 and 12, and the scale is smaller above the break, giving an emphasis to the
lower values. The graphs also lack an indication of the reliability of the
measurements and a statistical evaluation of the results.
View larger version (25K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 21A. —Examples of graphs in which changes in values for individual
patients are almost impossible to follow. A large amount of data ink was used.
(Reprinted with permission from
[19]) Graphs illustrate
changes before and after angioplasty in two vascular phenomena, ankle-brachial
pressure (A) and peak velocity (B). Discerning mean values
(thick dashed lines) is difficult. Limits for abnormal values
(thin dashed lines) are useful. Y-axis scaling for part
B is different below and above axis break, emphasizing lower values. No
indication of reliability or statistical tests for measurements are provided,
even for individual cases, so we cannot judge whether differences are
significant.
|
|
View larger version (23K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 21B. —Examples of graphs in which changes in values for individual
patients are almost impossible to follow. A large amount of data ink was used.
(Reprinted with permission from
[19]) Graphs illustrate
changes before and after angioplasty in two vascular phenomena, ankle-brachial
pressure (A) and peak velocity (B). Discerning mean values
(thick dashed lines) is difficult. Limits for abnormal values
(thin dashed lines) are useful. Y-axis scaling for part
B is different below and above axis break, emphasizing lower values. No
indication of reliability or statistical tests for measurements are provided,
even for individual cases, so we cannot judge whether differences are
significant.
|
|
The critique of the graphs in this article was designed to help the reader
understand the principles of good data presentation, in which economy,
clarity, and honesty are the essential guides.
Summary
Radiologists should apply to the selection and content of graphics
conveying radiologic data the same skills they use in the selection of
radiographic images for presentation or publication. This article has reviewed
the fundamentals for visual display of quantitative information from
radiologic studies. The truth about the data should be shown in an efficient
manner and the chartjunk minimized. Clarity and honesty are paramount.
Although meeting these criteria seems a valuable goal and an easy task to
accomplish, these examples of graphics from the recent literature suggest that
we need to scrutinize more carefully. Clarity of graphing leads to clarity of
thinking and of presentation.
References
- Tufte ER. The visual display of quantitative
information. Cheshire, CT: Graphics, 1983:51
-111
- Karlik SJ. Exploring and summarizing radiologic data.
AJR
2003;180:47
-54[Free Full Text]
- Izumi M, Hida A, Takagi Y, Kawabe Y, Eguchi K, Takashi N. MR
imaging of the salivary glands in Sicca syndrome: comparison of lipid profiles
and imaging in patients with hyperlipidemia and patients with Sjögren's
syndrome. AJR
2000;175:829
-834[Abstract/Free Full Text]
- Kuszyk BS, Bluemke DA, Choti MA, Horton KM, Magee CA, Fishman EK.
Contrast-enhanced CT of small hypovascular hepatic tumors: effect of lesion
enhancement on conspicuity in rabbits. AJR
2000;174:471
-475[Abstract/Free Full Text]
- Kozerke S, Hasenkam JM, Nygaard H, Paulsen PK, Pedersen EM,
Boesiger P. Heart-motion-adapted MR velocity mapping of blood velocity
distribution downstream of aortic valve prostheses: initial experience.
Radiology
2001;218:548
-555[Abstract/Free Full Text]
- Chernoff DM, Ritchie CJ, Higgins CB. Evaluation of electron beam CT
coronary angiography in healthy subjects. AJR
1997;169:93
-99[Abstract/Free Full Text]
- Harlow CL, Stears RLG, Zeligman BE, Archer DG. Diagnosis of bowel
obstruction on plain abdominal radiographs: significance of air-fluid levels
at different heights in the same loop of bowel. AJR
1993;161:291
-295[Abstract/Free Full Text]
- Ishigaki T, Endo T, Ikeda M, et al. Subtle pulmonary disease:
detection with computed radiography versus conventional chest radiography.
Radiology
1996;201:51
-60[Abstract/Free Full Text]
- Bosch JL, Tetteroo E, Mali WP, Hunik MGM. Iliac artery occlusive
disease: cost-effectiveness analysis of stent placement versus percutaneous
transluminal angioplasty. Radiology
1998;208:641
-648[Abstract/Free Full Text]
- Harrell GC, Floyd CE, Johnston GA, Ravin CE. Quality control
phantom for digital chest radiography. Radiology
1997;202:111
-116[Abstract/Free Full Text]
- McMahon H, Engelmann R, Behlen FM, et al. Computer-aided diagnosis
of pulmonary nodules: results of a large-scale observer test.
Radiology
1999;213:723
-726[Abstract/Free Full Text]
- Goldfarb LR, Alazraki NP, Eshima D, Eshima LA, Herda SC, Halkar RK.
Lymphoscintigraphic identification of sentinel lymph nodes: clinical
evaluation of 0.22 mm filtration of Tc-99m sulfur colloid.
Radiology
1998;208:505
-509[Abstract/Free Full Text]
- Minn H, Zasadny KR, Quint LE, Wall RL. Lung cancer: reproducibility
of quantitative measurements for evaluating 2-[F-18]-Fluoro-2-deoxy-d-glucose
uptake at PET. Radiology
1995;196:167
-173[Abstract/Free Full Text]
- Polak JF, Shemanski L, O'Leary DH, et al. Hypoechoic plaque at
ultrasound of the carotid artery: an independent risk factor for incident
stroke in adults age 65 years or older. Radiology
1998;208:649
-654[Abstract/Free Full Text]
- Szolar DH, Kammerhuter F. Quantitative CT evaluation of adrenal
gland masses: a step forward in the differentiation between adenomas and
non-adenomas? Radiology
1997;202:517
-521[Abstract/Free Full Text]
- Holhouser BA, Ashwal S, Luy GY, et al. Proton MR spectroscopy after
acute central nervous system injury: outcome prediction in neonates, infants
and children. Radiology
1997;202:487
-496[Abstract/Free Full Text]
- Chang Yue N, Arnold AM, Lonsteth WT, et al. Sulcal, ventricular and
white matter changes at MR imaging in the aging brain: data from the
cardiovascular health study. Radiology
1997;202:33
-37[Abstract/Free Full Text]
- Rubenstein JD, Kim JK, Henkelman RM. Effects of comparison and
recovery on bovine articular cartilage: appearance on MR images.
Radiology
1996;201:843
-850[Abstract/Free Full Text]
- Minar E, Pokrajac B, Ahmadi R, et al. Brachy-therapy for
prophylaxis of restenosis after long-segment femoropopliteal angioplasty:
pilot study. Radiology
1998;208:173
-179[Abstract/Free Full Text]
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
|
|
 
L. Joseph and C. Reinhold
Statistical Inference for Continuous Variables
Am. J. Roentgenol.,
April 1, 2005;
184(4):
1047 - 1056.
[Full Text]
[PDF]
|
|
|
Top
Introduction
Graphic Integrity
) for male
(A) and female (B) patients. Each patient data point represents
parotid gland size, age, and patient condition. Parotid gland size increased
in patients with hyperlipidemia (
) but not Sjögren's syndrome
(
). Mean values ± two standard deviations are plotted
(containing 95% of data) to provide visualization of spread of control data
versus patient values.
), tumor ([UNK]), aorta (