Designing a vascular segmentation "lookalike" task for non-medical
users.
For an overview of IO devices, possible interaction methods, and first
experiences with them, see
the section on IO devices.
Description of the task:
The task involves a vessel network (typically a vessel with at least one
bifurcation). A particular path through the vessel structure is automatically
chosen by the centerline algorithm. This usually determines the appropriate
vessel centerline correctly, so no user intervention is required. Then, the
diameter and shape of the vessel is automatically determined, but with
possible errors. Then, the user has to check the vessel segmentation
manually. If the vessel shape is incorrect, it has to be edited manually with
help of the scan data.
The task involves:
- analysis of a complex 3D structure.
- checking the diameter of each cylinder along the path according to the
pixel greyscale values and the overall 3D structure. Bifurcations will be the
most difficult and may require reverting to the full 3D view (rather than just
stretched view).
- switching between 3D interpretation and 2D contour editing (a precision
task).
We want to mimick the segmentation task as closely as possible.
However, can a segmentation-like 3D interpretation task be done by novices,
and, if so, is the experience of novices transferable to experience by
experts? In particular, experts are used to working with slice data, which
may be difficult to interpret by beginners but much easier with experience.
So, novices may prefer a 3D view while experts will not.
This means we cannot use the relative performance of slice vs 3D of our users
to predict the relative performance of medical experts. What we can
test with novices is how different kinds of novel representation and
navigation methods compare.
I propose a simple task that has some of the more interesting properties of
the segmentation task. This is a maze interpretation task. At its basis is a
traditional orthogonal rectangular maze (but in 3D, a cube maze). The idea is
that interpreting a 3D maze is roughly similar to vessel interpretation, but a
maze should be interpretable by any type of user. In particular, the maze
highlights interpretation of the overall vascular structure and its
relationship with individual bifurcations that need closer examination.
This type of maze has similarities to a vascular network, in that it is
basically a tree. Complex vascular networks look a lot like these 3D mazes.
For example:
The mazes I generate now look like this:
(click to enlarge)
Some example tasks:
-
A particular point or path in the maze is highlighted, and the user has to
name the number of bifurcations on one side of the point or along the path.
An alternative is to add artificial aberrations to the maze, and have the user
point out the aberrations.
The user may use the mouse to annotate segments, or clicks on the aberrations.
- A particular path is highlighted, and the user has to traverse along the
path, checking and correcting artificially induced errors at, in particular,
the junctions of the maze. This task follows the actual task in more detail,
in particular, it involves switching between 3D interpretation and the slice
editing task. A problem with this task is that any abstract,
beginner-friendly slice editing task will have dubious knowledge transfer to
expert users.
The "exhaustive search" nature of the task may be expressed by enabling a
"cursor" to slide along the path to be examined, for example using the cursor
keys. I am not sure whether existing medical software uses this, or only
enables a slice to be selected by clicking on a length projection (typically
the stretched view) with the mouse.
Research questions that we can readily address:
-
When does stereoscopic display have added value over other kinds of display?
Basically we can just test stereo 3D (using red/cyan glasses) against mono 3D.
- When is two-handed navigation better than mode-switch navigation? What
input device is best? There are several options here:
- Editing with mouse only, mouse buttons are used to switch between
navigation/edit
- Editing/navigation with 2 mice
- Editing with mouse, keyboard navigation
(wsadqe keys for rotate/roll for example)
- Editing with mouse, navigation with 3D joystick
- In what ways can I switch between different views? Basically there are at
least 3 views:
- 3D (volume) view: stereoscopic or not
- 2D projection (in particular, vessel slice view, but also other
projections)
- stretched view. This is a problematic one: I need more technical detail
about this view. For example: how is the nonlinear transformation performed,
and how does it deal with a vessel "doubling back" on itself? Does the view
only show voxels near the centerline?
- Does moving an object give extra 3D cues and do users use this to
interpret an object?
Other research questions which we want to do with experts:
- How can I edit a contour?
- regular 2D slice
- show the slices in the 3D view and edit them there, or in a zoom-in of the
3D view.
- Is navigation through the vessel by means of the cursor keys a good
option? Should the 3D view adapt automatically when navigating?
- How is best switched between contour editing and 3D interpretation?
- have the 3D and 2D views next to each other permanently or change the
screen layout?
- switch between input devices?
- switch between stereoscopic and mono image?
Considerations regarding input devices
With the basic maze task that I have now,
where the user interprets the model and has to select certain junctions, I
have the following two main interactions:
- select a segment in the maze
- rotate the model
It does not seem necessary to support model translation in addition to rotation,
since the model is small and does not require extensive navigation.
It seems best to do the selection wth the mouse. A 3D input device does not
seem attractive. The literature suggests that a 2D position device is best.
See also "selection using a one-eyed cursor in a fish tank VR environment".
For the combination with rotation I currently have the following set of "most
promising options":
- Single mouse: use the different mouse buttons in a traditional way (left
button=rotate, right button=select)
- Dual mouse: use a second mouse in the left hand for rotating (the regular
mouse in the right hand for selection).
- Trackball: use a trackball in the left hand (a trackball is analogous to
model rotation).
- Head tracker: use head position to rotate the model,
as described in the IR tracker experiments.
- Finger tracker: use the pointing finger of the left hand to rotate the
model. This can be done similar to the mouse, or using 3D position (how?).
See again the IR tracker experiments.
Here are some devices I decided to exclude from the list:
- Wiimote accelerometer. This accelerometer is not accurate enough to
determine roll/pitch, and my experience is that other accelerometers also have
accuracy problems. It would be interesting to find a very accurate one.
- Haptic device. The haptic device we have at Twente is accurate, but the
position translation and fatigue problems make it difficult to control.
- Magnetic trackers. It is well known that these trackers produce noisy
data and are not suitable for millimeter precision.
-