Comparison of navigation techniques for large digital images.

Medical images are examined on computer screens in a variety of contexts. Frequently, these images are larger than computer screens, and computer applications support different paradigms for user navigation of large images. The paper reports on a systematic investigation of what interaction techniques are the most effective for navigating images larger than the screen size for the purpose of detecting small image features. An experiment compares five different types of geometrically zoomable interaction techniques, each at two speeds (fast and slow update rates) for the task of finding a known feature in the image. There were statistically significant performance differences between several groupings of the techniques. The fast versions of the ArrowKey, Pointer, and ScrollBar performed the best. In general, techniques that enable both intuitive and systematic searching performed the best at the fast speed, while techniques that minimize the number of interactions with the image were more effective at the slow speed. Additionally, based on a postexperiment questionnaire and qualitative comparison, users expressed a clear preference for the Pointer technique, which allowed them to more freely and naturally interact with the image.

Introduction

Viewing images larger than the user’s display screen is now a common occurrence. It occurs both because the spatial resolution of digital images that people interact with continues to increase and because of the increasing variety of smaller resolution screens in use today (desktops, laptops, PDAs, cell phones, etc.). This leads to an increased need for interaction techniques that enable the user to successfully and quickly navigate images larger than their screen size.

People view large digital images on a computer screen in many different kinds of situations. This paper draws from work in many fields to address one of the most common tasks in medical imaging, finding a specific small-scale feature in a very large image. An example is mammographers looking for microcalcifications or masses in mammograms. For this study, large images are defined as images that have a spatial resolution significantly larger than their viewing device, i.e., at least several times larger in area. It may additionally be constrained by the user operating within a window on that screen that further constrains the available resolution. For instance, a user may wish to navigate a digital mammogram image that is 40,000 × 50,000 pixels on a personal computer screen that is 1,024 × 768 pixels in a window of size 800 × 600 pixels.

In the past, computer and network speeds limited the speed at which such large images could be manipulated by the display device, limiting the types of interaction techniques available and their effectiveness. As computer and network speeds have increased, it is now possible to interactively manipulate images by panning and zooming them in real time on most computer-based display systems, including the graphics cards found on standard personal computers. The availability of interactive techniques supporting real-time panning and zooming provides for the possibility of improved human–computer interactions. However, most interactions in existing commercial applications as well as freely available ones do not take advantage of improved interaction techniques or necessarily use the techniques best suited for capabilities of their particular display device. To test different interaction techniques, five different interaction techniques supported by imaging applications were selected.

In order to quantitatively compare the performance of different techniques, we must be able to measure their performance on a specific task. There are many types of tasks and contexts in which users view large images. In this study, we chose to examine the task of finding a particular small-scale feature within a large image. This task was chosen because it is a common task in medical imaging, as well as in other related fields such as satellite imaging.1,2 In addition to the interaction technique, the speed of updating the image view may affect the quality of the interaction. Several factors can affect the update rate, including processor speed and network connection speed. Increasingly, radiologists read from teleradiology systems, where images may be displayed on their local computer from a remote image server. To model this situation where images may be loaded over a slower internet connection, as compared to directly from the local computer memory, two display update rate conditions were tested. The slower update rate also corresponds to the typically slower computational speeds of small devices (PDAs, cell phones) and serves to model these situations as well. A change in the speed of image updates on the screen can dramatically affect the user experience resulting from the same interaction technique. To address this issue, we tested five different interaction techniques, with each technique evaluated with both a fast and a slow update rate.

Background and Related Work

There has been interest in viewing large digital images since the start of digital computers and especially since the advent of raster image displays. Several decades ago, researchers began to consider digital image interpretation in the context of image display.3 Today, digital image viewing and interpretation plays a vital role in many fields, including much of medical practice. Digital images are now routinely used for much of medical practice including radiology.4–6

This paper is concerned with navigational and diagnostic uses (as defined by Plaisant et al.7) of digital images when displayed on screens of significantly smaller size. We limited our focus to techniques used on standard computing devices, i.e., not having special displays or input devices and used geometric zooming. Nongeometrical methods (like fisheye lens zooming) are not considered because the size and spatial distortions that occur to the images are not acceptable in medical imaging practice. Interfaces that provide the ability to zoom and pan an image have been termed “zoomable” interfaces in the human–computer interaction literature.8 Two well-developed environments that support development and testing of general zoomable interfaces are the Pad++9 and Jazz toolkits.10 To date, few studies have examined digital image viewing from the perspective of maximizing effective interface design for the task of navigating and searching out features within a single large image. There is, however, a significant body of literature in related areas.

Studies on Related Topics

Many researchers have examined the transition from analog to digital presentations, especially in medical imaging.11–16 Substantial work has been done with nongeometrical zoomable interfaces including semantic zooming,8,17 distortion-based methods (fisheye),18–20 and sweet spots on large screens.21 A summary of these different types of methods can be found in Schaffer et al..22 Additionally, much work has focused on searching through collections of objects. Examples include a single image from a collection of images,9,23–26 viewing large text documents or collections of documents,22,27 and viewing web pages.28 Methods that involve changing the speed of panning depending on the zoom scale may have some relevance to our results. These methods have been developed to allow users to move slowly at small scales (fine detail) and more quickly over large scales (overviews). Cockburn et al.29 found that two different speed-dependent automatic zooming interfaces performed better than fixed speed or scrollbar interfaces when searching for notable locations in a large one-dimensional textual document. Ware and Fleet30 tested five different choices for automatically adjusting the panning speed, primarily based on zoom scale. They found that two of the adaptive automatic methods worked better than three other options, including fixed speed panning, for the task of finding small-scale boxes artificially added to a large map. Their task differs from our study in that their targets were easily identified at the fine-detail scale. Difficult-to-detect targets require slower, more careful panning at the fine-detail scale, which probably negates the advantage of automatic zooming methods for our task.

Closely Related Studies

One of the first articles addressing navigational techniques for large images was the article of Beard and Walker,31 which found that pointer-based pan and zoom techniques performed better than scrollbars for navigating large-image spaces to locate specific words located on tree nodes. They followed this work with a review of the requirements and design principles for radiological workstations32,33 and an evaluation of the relative effects of available screen space and system response time on the interpretation speed of radiologists.34,35 In general, faster response times for the user interface, larger screen space, and simpler interfaces (mental models) performed better.33 This was followed by timing studies that established that computer workstations using navigational techniques to interact with images larger than the physical screen size could perform as well or better than their analog radiology film-based displays.11,16,34,35 Gutwin and Fedak20 studied the effect of displaying standard workstation application interfaces on small screen devices like PDAs. They found that techniques that supported zooming (fisheye, standard zoom) were more effective than just panning and that determining which technique was most effective depended on the task. Kaptelinin36 studied scrollbars and pointer panning, the latter method evaluated with and without zooming and overviews. His test set was a large array of folder icons, with the overall image size nine times the screen size. Users were required to locate and open the folders to complete the task. He found the pointer panning technique performed faster than scrollbars and was qualitatively preferred, likely due to it not requiring panning movements to be broken down into separate horizontal and vertical scrollbar movements. Also, he found the addition of zooming to improve task speed. Hemminger37 evaluated several different digital large-image interaction techniques as a preliminary step in choosing one technique (Pointer) to compare computer monitor versus analog film display for mammography readings16. However, the evaluation was based on the users’ qualitative judgments and did not compare the techniques quantitatively.

Despite the relative lack of research in the specific area of digital-image-viewing techniques, many applications exist for viewing digital photographs, images, and maps. Online map providers such as Mapquest (available at http://www.mapquest.com, accessed September 2005) and Google Maps (available at http://maps.google.com/, accessed September 2005), as well as the National Imagery and Mapping Agency38 and the United States Geological Survey39 provide map viewing and navigating capabilities to site visitors. Specialized systems, such as the Senographe DMR (GE Medical Systems, Milwaukee, WI, USA), are used for detection tasks by radiologists; software packages such as ArcView GIS40 support digital viewing of feature (raster) data or image data. Berinstein41 reviewed five image-viewing software packages with zooming capabilities, VuePrint, VidFun, Lens, GraphX, and E-Z Viewer, which were frequently used by libraries. The transition from film to digital cameras for the consumer market has resulted in a wide selection of photographic image manipulation applications.

These tools use a variety of different interaction techniques to give viewers access to images at different resolutions. There are two basic classes of interactions involved. The first is zooming, which refers to the magnification of the image. The spatial resolution of the image as it is originally acquired is referred to as the “full resolution.” Different zoom levels that shrink the image in spatial resolution are provided so that the image can be shrunk down to fit the screen. The second operation is panning, which refers to the spatial movement through the image at its current zoom level. Most tools use some combination of these two techniques. Prominent paradigms for zooming in and out of images and some example applications that use them include: the use of onscreen buttons–toolbars,35–39 clicking within an image to magnify a small portion of that image (FFView available at http://www.feedface.com/projects/ffview.html, accessed September 2005), or clicking within the image to magnify the entire image with the clicked point at the center (ArcView GIS40). Prominent image-panning paradigms and example applications include the use of scroll bars (Mapquest available at http://www.mapquest.com, accessed September 2005; Microsoft Office Picture Manager and MicroSoft Office Paint available at http://microsoft.com, accessed September 2005; Adobe PhotoShop available at http://adobe.com/, 2005),40 moving a “magnification area” over the image in the manner of a magnifying glass (FFView available at http://www.feedface.com/projects/ffview.html, accessed September 2005), clicking on arrows or using the keyboard arrows to move over an image (Mapquest available at http://www.mapquest.com, accessed September 2005), panning vertically only via the mouse scroll wheel (Adobe PhotoShop available at http://adobe.com/, 2005),42 and dragging the image via a pointer device movement (Google Maps available at http://maps.google.com/, accessed September 2005; Microsoft Office Picture Manager and MicroSoft Office Paint available at http://microsoft.com, accessed September 2005).

Thus, while many systems exist to view digital images and digital image viewing is considered an important component of practice in many fields, there is no guidance from the literature regarding what geometric zoomable interaction techniques are best suited for navigating large images and, in particular, for the task of finding small features of interest within an image.

Materials and Methods

The main hypothesis was to determine which of five different commonly used types of interaction techniques were the most effective for helping observers detect small-scale features in large images and which of the techniques were qualitatively preferred by the users. Secondary aims include testing the main hypothesis when interaction techniques had slow update rates (such as might occur in teleradiology) and trying to identify major features of the interaction techniques that caused their success or failure. The study was comprised of both quantitative and qualitative parts. The quantitative part was the experiment to measure the users’ speed at finding features in large images when using different interaction techniques. There were three qualitative parts of the study: observations by the experimenter of the subjects during the experiment, a postexperiment questionnaire, and a qualitative comparison by the subject of all five interaction techniques on a single test image.

Pilot Experiment

To ensure we had developed the image-viewing techniques effectively and chosen appropriate targets within the images, we ran a pilot experiment. Three observers, who did not participate in the study, participated in the pilot. They each viewed 60 images using each of the five fast versions of the techniques to ensure that appropriate targets had been selected and to identify problems with the implementations of the techniques themselves. They then viewed ten images using each of the five slow versions of the techniques. Feedback from the pilot observers was used to refine the techniques and to eliminate target choices that, on average, were extremely simple or extremely difficult to locate. Measurements of the pilot observers completion times were also used to estimate the number of training trials needed to reach proficiency with the techniques. Once the experiment began, the techniques and targets were fixed.

Experimental Design

Quantitative

This study evaluated five different interaction techniques at two update rates (fast, slow) to determine which technique and update rate combinations were the most effective in terms of speed at finding a target within the image. Because the same interaction technique when used at a different update rate can have a substantially different user interaction, each of the combinations is treated as a separate method. An analysis of variance study design using a linear model for the task completion time was chosen to compare the performance of the ten different methods. The images used in the study were large grayscale satellite images with very small features to be detected. These images were chosen because they are of a similar size to the largest digital medical images; they were representative of the general visual task as well as the medical imaging specific task, and they allowed the use of student observers. In a prior work of Puff et al.,42 it was established that the student’s performance on such basic visual detection tasks served as a cost-effective surrogate for radiologist’s performance.

The task of finding a small target within a large image is naturally variable, affected by the image contents and each observer’s individual searching style. To minimize variance in each user’s performance, users received a significant amount of training to become proficient with the interaction method on which they would be tested. The number of study trials was also chosen to be large enough to help control for this variability. This led to having each user only perform with a single interaction method because the alternative (a within subject design) would have been prohibitive due to the number of trials required if each participant was to test with all ten interaction methods.

A total of 40 participants were recruited by flyers and e-mail for the study. Participants had to be over 18 years of age and have good vision (corrected was acceptable). They were students, faculty, and staff from the University of North Carolina at Chapel Hill (primarily graduate students from the School of Information and Library Science). Thirty-one participants were women and nine were men.

Each participant completed five demonstration images, 40 training images, and 120 study images for the experiment. They were each randomly assigned one of the ten interaction methods, which they used for the entire study. At the beginning of the first session, the participant completed an Institutional Review Board consent form. Then, the experimenter explained the purpose and format of the study and demonstrated the image-viewing tool with the five-image demonstration set. Next, the participant completed the training set of 40 images, followed by the study set. The study set consisted of 120 images in a randomized order, partitioned into four sets. The presentation order of the four image sets was counterbalanced across observers. Participants read images in multiple sessions. Most observers read in five separate sessions (training set and four study sets), although some completed it in fewer by doubling up sessions. Participants were required to take mandatory breaks (10 min/h) during the sessions to avoid fatigue. At the beginning of each new session, the participant was asked to complete a five-image retraining set to refamiliarize them with the interaction tool before beginning the next study image set. If time between sessions exceeded 1 week, participants were required to complete a ten-image retraining set.

Qualitative

During the experiment, the researcher took notes on the observer’s performance, problems they encountered, and unsolicited comments they made during the test. When participants had completed all of the image sets, they completed the postexperiment questionnaire (“Appendix 1”). Last, they were asked to try all of the interaction techniques using an additional test image to compare the methods and then rank them.

Images, Targets, and Screen Size

To test the viewing mechanisms, participants were asked to find targets, or specific details, within a number of digital grayscale photographs of Orange County, NC, USA. These photographs are 5,000 × 5,000 pixels in size and were produced by the US Geological Survey. Since participants were asked to find small details within the images, knowledge of Orange County did not assist participants in task completion. The targets were subparts of the full digital photograph and are 170 × 170 pixels in size. They were parts of small image features such landscapes, roads, and houses, which could be uniquely identified but only at high resolution. Target locations were evenly distributed across the images, so that results from participants who began each search in a particular location would not be biased. “Appendix 2” shows the distribution of targets within the images, for the 160 images in the training and test sets. The screen resolution of the computer display was 1,152 × 864 pixels, and the actual size of the display area for the image was 1,146 × 760 pixels. Thus, only about 3.5% of the full-resolution image could be shown on the screen at one time. “Appendix 3” shows a full image and an example target from that image.

Presentation and Zoom Levels

We tested five types of image-viewing techniques in the study. Each technique supported the following capabilities:

Ability to view both the image and the visual target at all times. The visual target was always onscreen at full resolution so that, if participants were viewing the image at full resolution, they would be able to see the target at an identical scale.

The entire image could be seen at once (by shrinking the image to fit the screen).

All parts of the image were able to be viewed at full resolution, although only a small portion of the full image could be seen at once when doing this.

Ability to choose a portion of the image as the target and get feedback as to whether the selection was correct or not.

An example screenshot is shown in Fig. 1, showing the Pointer interaction method at zoom level 3 (ZL3). The target can be seen in the upper-right corner.

Fig. 1

Sample screen from the Pointer interaction technique. The target is shown on the . The navigation overview is on the , with showing the current cursor location. The user is currently at Zoom Level 3 and positioned slightly above and left of the center of the full image.

Users would strike a key to begin the next trial. The application would time how long it took until they correctly identified the target. Identification of the target was done by the user hitting the spacebar while the cursor was over the target. Users would continue to search for and guess the target location until they found it correctly.

Four levels of zoom were defined to represent the image from a size where the whole image could be seen at once in ZL1 to the full-resolution image in ZL4. The choice of four zoom levels was determined by having the difference between adjacent zoom levels be a factor of 2 in each dimension based on previous work that found this to be an efficient ratio between zoom levels, performing faster than continuous zoom for similar tasks33,37. The image sizes for the four zoom levels were 675 × 675 pixels (ZL1), 1,250 × 1,250 pixels (ZL2), 2,500 × 2,500 pixels (ZL3), and 5,000 × 5,000 pixels (ZL4). Thus, when viewing the image at ZL4, only about 1/28th of the image could be seen on the screen at any one time. The MagLens and Section techniques used only one intermediate zoom level, in both cases similar to ZL3 of the other three techniques. The same terminology (ZL1, ZL2, ZL3, ZL4) is used to describe the zoom levels consistently between all the methods, with their specific differences described in the next section. “Appendix 4” contains an illustration of the four zoom levels. Resizing the image between zoom levels was done via a bilinear interpolation.

Interaction Techniques

Based on our review of the literature and techniques commonly available, we chose five different interaction techniques to evaluate.

ScrollBar

The ScrollBar technique allows the participant to pan around the picture by manipulating horizontal and vertical scroll bars at the right and bottom edges of the screen, similar to many current image and text viewing applications, in particular Microsoft Office applications. Zooming in and out of the image is accomplished using two onscreen buttons (ZoomIn and ZoomOut), located in the upper-left-hand corner of the screen. Four levels of zoom were supported. Image zooming is centered about the previous image center.

MagLens

The MagLens technique shows the entire image (ZL1) while providing a square area (512 × 512 pixels) that acts as a magnifying glass (showing a higher-resolution view underneath it). Using the left mouse button, the participant may pan the MagLens over the image to view all parts of the image at the current zoom level. Clicking the right mouse button dynamically changes the zoom level at which the area beneath the MagLens is viewed. Only three levels of zoom were supported (ZL1, ZL3, ZL4) because the incremental difference of using ZL2 for the MagLens area was not found to be effective in the pilot experiment and was eliminated. Thus, if the zoom level is set to ZL1 the participant is viewing the entire image at ZL1 with no part of the image zoomed in to see higher resolution. If the participant clicks once, the MagLens square would then show the image below it at ZL3 while the image outside of the MagLens stays at ZL1. Clicking again would increase the zoom of the MagLens area to ZL4, and a further click cycles back to ZL1 (no zoomed area). This interface style is found on generic image-processing applications, especially in the sciences, engineering, and medicine.

Pointer

The Pointer technique allows the participant to zoom in and out of the image by clicking the right (magnify) and left (minify) mouse buttons. Zooming is centered on the location of the pointing device (cursor on screen). Thus, the user can point to and zoom in directly on an area of interest as opposed to centering it first and then zooming. The Pointer method supports all four zoom levels. Panning is accomplished by holding the left mouse button down and dragging the cursor. We found that many users strongly identified with one of two mental models for the panning motion: either they were grabbing a viewer above the map and moving it, or they were grabbing the map and moving it below a fixed viewer. This corresponded to the movement of the mouse drag matching the movement of the view (a right drag caused rightward movement of the map) or the inverse (right drag caused leftward map movement), respectively. A software setting controlled this. The experimenter observed their initial reaction during the demonstration trials and configured the technique to their preferred mental model. The individual components (panning by dragging) and pointer-based zooming are often implemented, although this particular combined interface was not commonly available until recently (for instance it is now available in GoogleMaps (available at http://maps.google.com/, accessed November 2007) using the scrollwheel for continuous zoom). It is similar to the original Pad++ interface9 which used the center and right mouse buttons for zooming in and out. The Pointer interface used in this study is the same one qualitatively chosen as the best of these same five (fast) techniques in a medical imaging study by Hemminger.37

ArrowKey

The ArrowKey technique works similarly to the Pointer technique but uses the keyboard for manipulation instead of the mouse. The arrow keys on the keypad are used to pan the image in either a vertical or horizontal direction in small discrete steps. As with the Pointer interface, a software toggle controlled the correspondence between the key and the direction of movement and was configured to match the user’s preference. The ArrowKey method supported all four levels of zoom. Zooming is accomplished by clicking on the keypad Ins key (zoom in) or Del key (zoom out). The technique always zooms into and out of the image at the point that is at the center of the screen. This interface sometimes serves as a secondary interface to a pointer device for personal computer applications; it is more common as a primary interface on mobile devices which have only small keypads for input.

Section

This technique conceptually divides each image into equal size sections and provides direct access to each section through the single push of a key. A section of keys on the computer keyboard were mapped to the image sections so as to maintain a spatial correspondence, i.e., pushing the key in the upper right causes the upper-right section of the image to be shown at a higher resolution. In our experiment, the screen area was divided into nine rectangles, which were mapped to the one to nine buttons on the keyboard’s numeric keypad. The upper-left-hand section of the image would be selected and displayed at ZL3 by hitting key 7, the upper center by key 8, the upper right by key 9, and so forth. Once zoomed in to ZL3, the participant may zoom in further to ZL4 to see a portion of the ZL3 image at full resolution by striking another one of the one to nine keys. Thus, this technique allows the participant to view a total of 81 separate full-resolution sections, all accessible by two keystrokes. For instance, to see the upper rightmost of 81 sections, the participant would hit key 9 followed by key 9. To zoom out of any section, the participant presses the ZoomOut (insert) key on the numeric keypad. An overlap of the sections is intentionally built in at the section boundaries, as illustrated in “Appendix 5.” This allows participants to access targets that may otherwise have been split across section boundaries. The Section method supports three levels of zoom (ZL1, ZL3, and ZL4) similar to MagLens because the pilot experiment found the use of ZL2 to be a detriment for this technique. This interaction is sometimes implemented with fewer sections (for example quadrant-based zooming). It is less common than the other choices and probably more suited to mobile devices that have numeric keypads but not attached pointing devices.

Navigation Overview

Many systems provide a separate navigation window showing the user what portion of the entire image they are currently viewing7,43. In our work evaluating several zoomable interfaces for medical image display37, we found that, when the zooming interactions operated in real time and the full image could be accessed in less than 1 s (for instance via two mouse clicks or two keystrokes), users preferred to operate directly on the image instead of looking to a separate navigation view. Hornbaek et al.44 reported similar findings for an interface with a larger number of incremental zoom levels (20). They found that users actually performed faster without the navigation view and switching between the navigation and the detail view used more time and added complexity to the task. Because some of the techniques tested in this study (particularly the slow update rate ones) might not perform as well without a navigation view, a navigation window (100 × 100 pixels in the upper-left corner) was included as part of all of the techniques. Based on the pilot study and guidelines7,31,44–46 established for navigation overview windows, the overview window was constructed so that it was tightly coupled to the detail window, showed the current location of the cursor, and kept small to leave as much of the screen real estate for the detail window as possible, which was crucial for this study’s task.

We developed ten viewing tools corresponding to the ten methods and implemented them as Java 2.0 programs, running on a Dell 8200 computer with 1 GB of memory, and a 20-in. color Sony Trinitron cathode ray tube monitor. The viewing tools, an example image and instructions, are available at http://ils.unc.edu/bmh/pubs/PanZoom/.

Results

Quantitative

We analyzed the training (first 40 images) and test images (numbered 41–160) to see if the observers reached asymptote performance with their interaction method by the end of their training, so that their test results would not be biased by observers continuing to significantly improve during the study trials. Time for each subject was modeled using least squares as a function of trial number with a modified Michaelis–Menten function which is nonlinear, monotonic, and decreasing to an asymptote. All observers reached asymptote performance by the end of training with most achieving it within the first 10–15 training cases. An example observer’s reading times with asymptote curve fit is seen in “Appendix 6.”

The primary quantitative analysis was to compare the ten different methods (five techniques each at two speeds) based on how quickly observers could complete the feature-finding task using that method. Table 1 summarizes the mean time and standard deviation for each method, calculated across all observers and all trials. To determine whether a particular method performed faster than another, the mean task completion times were compared using the SAS (Cary, NC, USA) GENMOD repeated-measures regression test (1 df, complete analysis in “Appendix 7”). A P value of 0.05 or less indicates the null hypothesis that the techniques have the same performance, which is rejected, and that the performance of the two techniques is statistically significantly different from each other. Using the results from this analysis, we grouped the methods into performance groups. Table 1 shows the mean task completion times in seconds averaged across all observers for each method and the performance groupings. Methods were placed in the same performance group if they had similar mean times and did not have statistically significant differences in mean times from all other members of the performance group (using SAS GLM Tukey’s Studentized Range Test). This segregated the methods into four groups (Table 1). Part of the reason for grouping the techniques is that the group rankings are probably more informative than the individual rank ordering of methods, due to the large standard deviations in detection times due to image and observer effects, as seen in Table 1. A further regression analysis was conducted to compare these resulting groups. All of the groups were found to be statistically significantly different from one another (p value < 0.05), with the exception of group 1 versus group 2. A power analysis based on the existing data show that the study would have to increase from four to seven observers per method in order to reduce the variance sufficiently to demonstrate the difference between group 1 and group 2 at a statistically significantly level.

Table 1.

The Mean Task Completion Times (seconds)

Performance group	Method name	Mean time	SD
1	ArrowKey Fast	76	76
1	Pointer Fast	79	100
1	ScrollBar Fast	84	108
2	Section Fast	97	127
2	Section Slow	97	131
2	Scrollbar Slow	98	94
3	MagLens Slow	117	165
3	ArrowKey Slow	119	134
3	Pointer Slow	128	146
4	MagLens Fast	155	176

A regression analysis was also performed to examine the significance of the other two factors (observer and image). The largest determining factor was the method, with the observer and image effect each approximately one third the magnitude. Table 2 shows how much each of the main effects contributes to determining the speed of detecting targets.

Table 2

Main Effects

Factor	df	F value	Pr > F
Method	0	21.96	<0.0001
Observer	30	8.76	<0.0001
Image	115	9.42	<0.0001

The last analysis determined whether the slow versions of techniques generally performed the same or differently than the fast versions of the techniques. A comparison of differences in mean task completion times between the fast and slow versions of each of the five techniques to zero (SAS GENMOD analysis, 5 df) determined that the fast techniques were statistically significantly different from the slow ones (P value of 0.047). It is evident that they are faster from Table 1, with the exception of the MagLens fast technique which observers had some difficulties with, resulting in it being the poorest performer.

A significant amount of valuable information resulted from observing the participants, from the survey, and from the postexperiment testing. We summarize only the highlights here but have included much of the rich qualitative details in “Appendix 8.”

Our observations of the observers closely matched both their comments and their rankings of the techniques. Table 3 shows the rankings of the interaction techniques by the observers, based on their trying each of the techniques at the conclusion of the study. Observers assigned the techniques rankings of 1–5 (1 being the best). The Pointer technique was listed by almost all observers as the best technique. The rest of the techniques all clustered at slightly below average.

Table 3

Observers’ Rankings of Techniques on a 1–5 Scale

Technique	Average ranking
Pointer	1.77
MagLens	3.12
ArrowKey	3.28
ScrollBar	3.30
Section	3.51

Reasons the observers gave for favoring the Pointer method was the natural control it gave them in panning around the image, precise control of the zooming, maintaining context (location in the overall image), and speed of operations. The ArrowKey method was also favored for its speed and precise control of panning and zooming. Participants did not rank it as high because they found the panning motion to be “less smooth” and it was “harder to scan” than with the Pointer method. They did find the ArrowKey technique very effective for systematic searching. Some users found the MagLens interaction desirable because you always maintained the context of where you were in the overall full-resolution image. It was also considered to be a more familiar paradigm than some of the other techniques like the Section. However, many users felt it was difficult to use in practice, saying it was “hard on the eyes” and “is a pain” and several observers who used it complained that it was disorienting to use, with one becoming dizzy as a result. The ScrollBar technique was considered “familiar” yet “old-fashioned.” Users felt it gave them good control but with too limited flexibility (i.e., only being able to pan in one dimension at a time versus two for most of the other techniques). Only two of the eight participants who had used the technique in the study ranked it in their top two choices. The Section technique was the least favored of all the techniques. Panning of the image is not directly supported by this technique, in that users have to step up a zoom level and then back down again in an adjacent section to effect a “pan” operation. Users felt this did not allow a natural panning exploration to occur, that too many button clicks were required to pan around, and that the constant zooming in and out frequently caused a loss of context.

The navigation view was very rarely used except for experimenting with it in training. The few instances where it was observed being used during the test cases were in the slow versions.

Discussion

Our results indicate that some interaction techniques perform quantitatively better for feature detection types of tasks. Integrating the results from the quantitative and qualitative portions of the study did yield several consistent overall themes, and a clearer understanding of the benefits and shortcomings of the individual techniques is presented in this section. It is important to remember, though, that the performance of interaction techniques will clearly depend on the task, and these results may not hold for other types of tasks. Additionally, the chosen surrogate visual detection task is not representative of all types of medical imaging tasks.

Overall Themes

Intuitive and Easy-to-Use Interface Favored From the qualitative feedback, users expressed clear preferences for intuitive, easy-to-use, and highly interactive user interface techniques. There were common elements to the techniques that performed well quantitatively and were preferred qualitatively. The top three performing techniques supported natural and easy ways to perform image panning. They supported both systematic and intuitive target searching. The most preferred method, Pointer, was favored in a large part because it had the most natural interaction for panning, with hand motion of the pointer corresponding to moving the image viewpoint. The most preferred methods (Pointer, ArrowKey) supported easy control of zooming, in that zoom levels could be selected without the observer moving their hand. Techniques that had more challenging mental models (Section) or difficult interactions (MagLens) were not favored and did not perform as well.

Simple Interface Favored Techniques that minimized interactions (keystrokes, mouse clicks, hand motions) tended to perform better, as might be predicted by Goals, Operators, Methods, and Selection47 modeling of the techniques. The Pointer and ArrowKey had the most efficient interactions due to the hand remaining on input device (mouse or arrow keys, respectively), and only one interaction (click) is required for both pan and zoom operations. The Scrollbar method was perhaps the least efficient due to having to move the pointer between three areas and click on small controls (vertical and horizontal scrollbars and the zoom buttons). This was reflected in the user’s comments and rankings which made it clear that they did not favor this technique because it did not support natural and quick panning and was too cumbersome for more generalized tasks. However, the Scrollbar method performed well quantitatively for the feature detection task because all the users of this technique adopted a systemic way to scan the image (they scrolled across the image a “row” at a time using only one scrollbar control). Additionally, using multimodal interfaces may add mental distractions for the user. It is possible that the MagLens and Scrollbar interactions may have suffered from this because these two interactions utilized both the mouse and keyboard while the other techniques were primarily keyboard-based (Section, ArrowKey) or mouse-based (Pointer).

Faster and Real-Time Interactions Preferred Users clearly favored the faster update rate versions of techniques and also performed better with them in all cases except the MagLens technique, where the fast version had worse performance likely due to the users losing context and getting confused about what part of the image they had already viewed.

Individual Techniques

ArrowKeys This was one of the top performers and, while it was significantly behind the Pointer technique in user preference, it was generally favorably reviewed by observers. While this technique was not as natural as the mouse-panning interaction of the Pointer technique, the small discrete movements (left, right, up, down) were easily understood and utilized by the observers. As with the Pointer method, the slow version of this technique did not perform as well because of the reduced interactivity of the pan operation.

MagLens While this technique was familiar to most users, and favored by some, it was generally not preferred by those who used it in the experiment, and it performed the worst overall of all the techniques. Interestingly, the fast version was by far the slowest in performance. Users of the fast version tended to try to interactively pan more. When they did this, they lost their position (context) and often became disoriented with respect to what territory they had covered already. The users of the slower version tended to adopt a more methodical search pattern for covering the image at a high zoom level and ended up being more efficient.

Pointer This was one of the top performing techniques and the clear favorite of the observers. The interface lends itself well both to systemic tasks like the feature detection task of this experiment and more general tasks, such as manipulating large images or following map driving directions. As computer and graphics card speeds have increased, the panning part of the Pointer interaction (dragging the mouse) is becoming fairly common, and having interactive zoom is beginning to appear in tools. Currently, most tools have a separate interaction for zooming, as in MapQuest, which zooms by mouse clicks on a scale on the screen or keystrokes on the keyboard. This is less efficient than having both the zoom and the panning operations accomplished from the pointing device.33 An easy way to do this is to zoom via the scrollwheel now commonly found on mouse devices, and this has been adopted by recent applications (for instance GoogleMaps42 now supports this). This technique is strongly dependent on a fast interaction. The natural connection between the panning motion of the mouse and the movement of the image on the screen was lost due to the update delay in the slow version of the pointer interaction. The result was that the slow version was not favored by users and was next to last in performance.

Scrollbar The Scrollbar method was familiar to users. They found it satisfactory for one-dimensional scrolling, as is commonly found in text viewers. However, it was generally viewed as cumbersome for navigating in two dimensions because of having to separately manipulate the vertical and horizontal scrollbar controls. In this experiment, users were able to adapt the task to a series of systematic searches along “rows” of the image, reducing their usage to manipulating a single scrollbar control to move across one “row” at a time. This allowed them to perform efficiently with both the fast and the slow versions of the technique.

Section The Section method was the least favored by the observers because most were not familiar with the technique, and the mental model was not as natural to them. However, users were able to become efficient with this technique, and both the fast and the slow version were in the top five in performance. It appeared that the slow version performed as well as the fast version because users tended to not rely on many quick panning motions but instead adopted a systematic section-by-section search pattern, which was not significantly affected by the difference in the slow and fast update rates.

This experiment dealt with a particular feature detection task, and given sufficient training users were, in most cases, able to adapt to the technique they utilized to efficiently perform the task. For most of the techniques, this resulted in the users scanning out the image in rows, with the height of the row being the size of the image seen at either ZL3 or ZL4 (depending on user preference). This type of serialized scanning interaction is formalized in several disciplines; for instance, it was popularized by Laszlo Tabar as a method of training radiology residents in detecting microcalcifications in mammography. The ArrowKey, Scrollbar, and Section techniques support this type of highly structured, linear movement in vertical or horizontal directions, especially well. They are less well suited to supporting navigation in two dimensions, such as following natural objects or anatomy. Observers commented that the Pointer method seemed much more effective for these types of interactions as well as for more general-purpose navigation.

Several factors affect the choice of the technique to utilize in a given situation. In addition to the task, the update rate of the display device and the types of interactions supported by the display device (keyboard only, cell phone–PDA keypad only) are key factors. For devices such as personal computers that commonly have pointing devices and fast update rates, the Pointer method would likely be an effective choice across a wide range of applications. If the update rate is not fast, then a different technique than the Pointer method may be more optimal (e.g., Section or Scrollbar). The ArrowKey and Section interfaces do not require a pointing device and thus may be better suited for small mobile devices such as cell phones and PDAs.

Since the fast versions of the techniques performed significantly better than the slow versions, there is not a single technique that can be considered the best choice for working well under both update conditions. Thus, applications that may be used under both conditions should consider offering more than one interface technique to the user. For this particular task, if only a single technique could be supported, then the Section and Scrollbar techniques might be good candidates since both the slow and fast versions of these techniques were in the top two performance groups.

Table 4

Contrast Estimate Results

Label	Estimate	Standard error	Alpha	Confidence limits		Chi-square	Pr > Chi sq
1 vs 2	5.4146	11.4895	0.05	−17.1044	27.9335	0.22	0.6375
1 vs 3	−70.8000	13.2604	0.05	−96.7900	−44.8100	28.51	<0.0001
1 vs 4	7.9250	17.9412	0.05	−27.2390	43.0890	0.20	0.6587
1 vs 5	−13.0771	10.3176	0.05	−33.2991	7.1450	1.61	0.2050
1 vs 6	−13.9833	12.9151	0.05	−39.2966	11.3299	1.17	0.2789
1 vs 7	−44.0583	16.3799	0.05	−76.1623	−11.9544	7.23	0.0071
1 vs 8	−33.1125	12.8421	0.05	−58.2826	−7.9424	6.65	0.0099
1 vs 9	−34.9500	15.9704	0.05	−66.2513	−3.6487	4.79	0.0286
1 vs 10	−12.3521	25.9969	0.05	−63.3052	38.6010	0.23	0.6347
2 vs 3	−76.2146	14.1289	0.05	−103.907	−48.5224	29.10	<0.0001
2 vs 4	2.5104	18.5923	0.05	−33.9298	38.9506	0.02	0.8926
2 vs 5	−18.4917	11.4123	0.05	−40.8593	3.8759	2.63	0.1052
2 vs 6	−19.3979	13.8054	0.05	−46.4560	7.6601	1.97	0.1600
2 vs 7	−49.4729	17.0906	0.05	−82.9698	−15.9760	8.38	0.0038
2 vs 8	−38.5271	13.7371	0.05	−65.4513	−11.6029	7.87	0.0050
2 vs 9	−40.3646	16.6985	0.05	−73.0930	−7.6361	5.84	0.0156
2 vs 10	−17.7667	26.4505	0.05	−69.6087	34.0753	0.45	0.5018
3 vs 4	78.7250	19.7358	0.05	40.0434	117.4066	15.91	<0.0001
3 vs 5	57.7229	13.1936	0.05	31.8640	83.5818	19.14	<0.0001
3 vs 6	56.8167	15.3108	0.05	26.8081	86.8252	13.77	0.0002
3 vs 7	26.7417	18.3281	0.05	−9.1807	62.6640	2.13	0.1446
3 vs 8	37.6875	15.2492	0.05	7.7996	67.5754	6.11	0.0135
3 vs 9	35.8500	17.9630	0.05	0.6431	71.0569	3.98	0.0460
3 vs 10	58.4479	27.2665	0.05	5.0067	111.8892	4.59	0.0321
4 vs 5	−21.0021	17.8918	0.05	−56.0694	14.0652	1.38	0.2405
4 vs 6	−21.9083	19.5055	0.05	−60.1385	16.3218	1.26	0.2614
4 vs 7	−51.9833	21.9537	0.05	−95.0118	−8.9549	5.61	0.0179
4 vs 8	−41.0375	19.4573	0.05	−79.1730	−2.9020	4.45	0.0349
4 vs 9	−42.8750	21.6499	0.05	−85.3080	−0.4420	3.92	0.0477
4 vs 10	−20.2771	29.8246	0.05	−78.7322	38.1780	0.46	0.4966
5 vs 6	−0.9063	12.8465	0.05	−26.0849	24.2724	0.00	0.9438
5 vs 7	−30.9813	16.3258	0.05	−62.9792	1.0167	3.60	0.0577
5 vs 8	−20.0354	12.7731	0.05	−45.0702	4.9993	2.46	0.1167
5 vs 9	−21.8729	15.9149	0.05	−53.0655	9.3197	1.89	0.1693
5 vs 10	0.7250	25.9629	0.05	−50.1614	51.6114	0.00	0.9777
6 vs 7	−30.0750	18.0798	0.05	−65.5108	5.3608	2.77	0.0962
6 vs 8	−19.1292	14.9499	0.05	−48.4305	10.1722	1.64	0.2007
6 vs 9	−20.9667	17.7097	0.05	−55.6770	13.7437	1.40	0.2364
6 vs 10	1.6313	27.1002	0.05	−51.4842	54.7467	0.00	0.9520
7 vs 8	10.9458	18.0277	0.05	−24.3879	46.2796	0.37	0.5437
7 vs 9	9.1083	20.3748	0.05	−30.8255	49.0422	0.20	0.6548
7 vs 10	31.7063	28.9123	0.05	−24.9608	88.3733	1.20	0.2728
8 vs 9	-1.8375	17.6565	0.05	−36.4436	32.7686	0.01	0.9171
8 vs 10	20.7604	27.0655	0.05	−32.2870	73.8078	0.59	0.4431
9 vs 10	22.5979	28.6823	0.05	−33.6183	78.8141	0.62	0.4308

The first column “label” in the statistical analysis above refers to pair-wise comparison of two methods. The methods are referred to by their numeric code; the coding scheme is shown below.