8. Research: “The influence of automation on tumor contouring”
Aselmaa, A., van Herk, M., Song, Y. et al. The influence of automation on tumor contouring. Cogn Tech Work 19, 795–808 (2017). https://doi.org/10.1007/s10111-017-
0436-0
20. General
How much the patient
deviates from the average
human in the region of
interest.
The quality of the CT scan
Prosthetics (dental fillings).
Placement of the tumor.
Doctor Specific
Doctors experience.
Doctors training
Doctors personal bias
Doctors specialty
AI specific
Size of the Training Dataset
Quality of the Training
Dataset
Context Factors
24. Uncertainty indicator
Translates the layer
uncertainty of the
contours into a scalar
value (percentage).
AI aided manipulations
By indicating where the
doctor does not agree
with the AI contouring the
AI can reevaluate the
choices and propose
changes.
VR environment
Use VR to view the model
as a whole and make
changes directly to the
model instead of on a per-
slice basis.
User Interface
25. Uncertainty indicator
Translates the layer
uncertainty of the
contours into a scalar
value (percentage).
AI aided manipulations
By indicating where the
doctor does not agree
with the AI contouring the
AI can reevaluate the
choices and propose
changes.
VR environment
Use VR to view the model
as a whole and make
changes directly to the
model instead of on a per-
slice basis.
User Interface
27. The use of colour is used to
indicate different tissue.
Green is nerve, violet is
gland, and orange is bone.
Refined Uncertainty indicator
28. By clicking on the uncertainty
collum the user can order the
organs on their uncertainty in
ascending or descending
order.
Refined Uncertainty indicator
29. There is a mistake in the contours that needs to be corrected.
AI Assisted Contouring
30. This symbol indicates on which layers the high uncertainty occurs
AI Assisted Contouring
31. The RTT starts to adjust the contours.
AI Assisted Contouring
32. Based on the changes the RTT made, RYCA can re-evaluate the contours.
AI Assisted Contouring
Then, compared it to the AI-infused workflow, that the RTT is removed from the flow and the whole process takes less time.
We wanted to see if the current workflows can give us some insight to start with, therefore by overlapping the two workflows (with and without AI), we could compare some important information.
We figured out that there are some tasks that AI can do precisely with a low uncertainty level.
However, RTT might be too slow and uncertain in those tasks.
On the other hand, there are some tasks that RTT does with higher certainty level than AI.
So at least one of them (AI or RTT) is good and efficient in those tasks.
The most problematic tasks are those that neither AI not RTT can do it with high certainty.
Based on the previous slide, our first attempt to make an overlapped workflow that is both fast and consistent, was to implement RTT as a peer reviewer in tasks that AI is weak at. If RTT was also uncertain, RO comes to triple check the result.
Therefore, we would reduce the pressure on RO.
Figure 1 Contouring workflow with RTT (2A) and AI (2B)
Compared contouring with RTT and AI for 3 different patient cases
Figure 2 (based on Figure 1) AI seems to be more time efficient
(C-1 is case 1, C-2 is case 2 and C-3 is case 3, All is all of them together)
https://link.springer.com/article/10.1007/s10111-017-0436-0
Other research:Tumor delineation: The weakest link in the search for accuracy in radiotherapy
It can be subjective and observer-dependent.
Some researchers have identified the lack of continuous education and training as a cause of the variability in tumor delineation
It is also recommended that radiation oncologists should collaborate with other specialties, such as radiologists.
https://www.jmp.org.in/article.asp?issn=0971-6203;year=2008;volume=33;issue=4;spage=136;epage=140;aulast=Njeh
Other research: applications and limitations of machine learning in radiotherapy
high degree of interobserver variability
Atlas method is the most commonly used, however
atlas-based methods are highly sensitive to the atlas-selection strategy,48 as well as the robustness of the—often time-consuming—registration itself
https://www.birpublications.org/doi/epub/10.1259/bjr.20190001
We brainstormed all the possibilities with the aim of reducing contouring time while enhancing accuracy.
The following slides demonstrate 5 idea proposals.
P.S. Could the RTT learn from their mistakes as well as the AI? Yes! We implemented this comment in slide 23.
Why is there no feedback to the RTT so they can learn too, like the AI?
We needed more information to evaluate and choose between those ideas.
Hence, we studied the context and the goal of this process.
We divided the goals in two parts, Primary and Secondary goals that you can see in this slide.
Activity Theory is a theoretical framework for the analysis and understanding of human interaction through their use of tools and artefacts
Based on activity theory, the Motives, Goals, and Subconcious are described.
And lastly, the contextual factors are described.
Q: Is giving the RTT the entire organ for context the only way?
A: we think this is the best way to be more accurate. However, they know the organ they are working on has a high uncertainty level.
Based on the reflections during presentation, we liked the idea that RTT can also learn their mistakes from RO refinements, so we implemented it in the workflow.
Uncertainty indicator: the uncertainty percentage is used to forward the organ for reviewing to either the RTT first and then the RO if it is highly uncertain. Or to the RO directly if the uncertainty is low.
We chose to continue with the Uncertainty indicator and the AI aided manipulations but not the VR environment.
The idea behind the VR environment was to cut down the time spend on misc tasks such as between slide navigation. However, to say that this new interaction would be faster while maintaining the same precision would be a huge assumption so we chose to not use it.
This is just a selected part of the UI for presentation relevance. This part of the UI is used to navigate between organs, patients and different scans.
Q: Why do you use uncertainty rather than certainty?
A: We chose to indicate uncertainty because that indicates how much initial outlines is left to be done, and how much more the RTT’s opinion supersedes that of RYCA. Moreover, we aim to reduce uncertainty level more and more every time that RYCA learns.
Q: The RTT still has to scroll through all the slices in order to find the uncertainty within the organ.
A: We added an indicator (as seen in slide 30) that shows on which slices the high uncertainty occurs.
Q: Why are all the colours green?
A: this is how we got the data and assumed the colours should be used to indicate what kind of material the organ is (for instance bone, tumor, …) So we did not find it proper to change colors for uncertainty level when we don’t know the mental background of the color indications for RTT and RO. We also implemented this in slide 27-28
This is just a selected part of the UI for presentation relevance visible on the same screen as the previous part of the UI. The UI for navigating the slices is kept mostly the same. The two added functions are the Uncertainty indicator of the per slice basis as shown in the next slide.
This indicator symbol is added to aid the RTT/ RO in finding where the higher uncertainties are, in order to correct them and decrease the time it takes to find them (misc/between slice navigation).
We chose to use a pop up to get the attention of the user. If it would happen in the corner of the screen the user would not notice or continually check if the AI proposes new contours, thereby increasing mental load.The placement of the Popup should be in the middle because that area is mostly not occupied with critical information (such as the center of CT scans or the 3D model). The popup should be a bit smaller so it does not block sight of the CT scans, but we chose to make it a little bit bigger for this presentation.
After the AI changed the model, it is now more certain that this is correct due to the input of the RTT/RO so it will remove the markers.