Transcript
Well, they are my disclosures, I will just like to tell you that Brainlab is in the development of this software. It uses big data and even uses deep learning for making it better than this version. I evaluated this.
I think you all know this, like sitting in front of your desk and doing some slice-wise, contouring of organs-at-risk. And in our clinic, yeah, delineation of organs-at-risk is a task for the young physicians. And we think, or I think, it's time consuming and repetitive. Actually, we have the time for...you know, we have the chance to alternate tasks which are repetitive and Brainlab says we have a tool for automatic segmentation and I were thinking, "Yeah, okay. But how good is it?" Sorry for the German 't' in it.
What I do, I use the Spine SRS Contouring Workflow in version 1.0 and just uses the Anatomic mapping tool. This tool just gives tissue classes after importing a CT data set. After that, it does some registrations to a predefined ethos model to just generate the organs-at-risk, like you can see here, it's lungs, heart, liver, and it even segmentates all vertebra.
For, yeah, my evaluation now, I used some data sets from Kiel. It was 29 patients. We just treated with indication of spinal metastasis in the beginning of 2017. And with these patient, which were a bit mixtured because we have, like, big CTVs. We have 33% of SBRT and 42% patients with like normal CTVs of 2 to 4 vertebra, which were all delineated with a, yeah, manually and contouring platform with variant.
We exported the OAR structure sets and compared them with the automatically generated structures from Brainlab Elements. These were the organs-at-risk we evaluated and for this [inaudible 00:02:09.701], we took two quantitative measurement. The first is the Dice Similarity Coefficient as some of you will know. It's just like equivitent which, yeah, tested the overlap of two or three dimensional structures. And it ranges from 0, no overlap, to 1, total overlap.
The second, or the Hausdorff Distance, which is a little bit more complicated because it takes of distant points from the two surfaces of the structures and says how big is its distance.
When you see this on box plots, it a bit looks like that the dices were just in the range between lot 0.7 and 1, except for the esophagus, the box plots. The Hausdorff Distance is, yeah, far more complicated to interpret. I just wanna turn out the spinal cord because here we see a problem with, like, small data. Because in our clinic, the spinal cord is contoured not only until L1, L2 how the software does it. But if we are just far below in the area, we just use spinal cord and call the [inaudible 00:03:18.170] as one structure so we have here a problem with the surface points. So I would say there's a...it's complicated to just use this quantitative measurement because for a clinical routine, the structures could have been good as well.
That's why we added a kind of qualitative review where we give every structure with automatically generated score about 1, 2, or 3. And 1 means it's useful without corrections, 2 means it's useful with minor corrections, and 3 means the structure is not useful. Here, have it for the heart. And it looks like the score of 1 correlates with High Dices and the score of 2 even has some High Dices but especially, we have still, maybe the dice of around 0.4 and we can say we can accept it in clinical routine. We've done this for all the structures and on this slide, I just wanna point out that even if we have some problems with, like, the lungs, it was not useful but the dice was still high.
So just to make a decision how good this software is, you have to look at all the quantitative and the qualitative measurements. And I have still for spinal couture here and esophagus. And I would just say in the end that for most of the structures, this software is ready and useful in clinical routine, except for the esophagus where it is not working. And actually, I can add, the version 1.5 is a bit better. Well, I'm still looking forward when we can alternate this, yeah, segmentation process in our clinical routine. Thank you.
I think you all know this, like sitting in front of your desk and doing some slice-wise, contouring of organs-at-risk. And in our clinic, yeah, delineation of organs-at-risk is a task for the young physicians. And we think, or I think, it's time consuming and repetitive. Actually, we have the time for...you know, we have the chance to alternate tasks which are repetitive and Brainlab says we have a tool for automatic segmentation and I were thinking, "Yeah, okay. But how good is it?" Sorry for the German 't' in it.
What I do, I use the Spine SRS Contouring Workflow in version 1.0 and just uses the Anatomic mapping tool. This tool just gives tissue classes after importing a CT data set. After that, it does some registrations to a predefined ethos model to just generate the organs-at-risk, like you can see here, it's lungs, heart, liver, and it even segmentates all vertebra.
For, yeah, my evaluation now, I used some data sets from Kiel. It was 29 patients. We just treated with indication of spinal metastasis in the beginning of 2017. And with these patient, which were a bit mixtured because we have, like, big CTVs. We have 33% of SBRT and 42% patients with like normal CTVs of 2 to 4 vertebra, which were all delineated with a, yeah, manually and contouring platform with variant.
We exported the OAR structure sets and compared them with the automatically generated structures from Brainlab Elements. These were the organs-at-risk we evaluated and for this [inaudible 00:02:09.701], we took two quantitative measurement. The first is the Dice Similarity Coefficient as some of you will know. It's just like equivitent which, yeah, tested the overlap of two or three dimensional structures. And it ranges from 0, no overlap, to 1, total overlap.
The second, or the Hausdorff Distance, which is a little bit more complicated because it takes of distant points from the two surfaces of the structures and says how big is its distance.
When you see this on box plots, it a bit looks like that the dices were just in the range between lot 0.7 and 1, except for the esophagus, the box plots. The Hausdorff Distance is, yeah, far more complicated to interpret. I just wanna turn out the spinal cord because here we see a problem with, like, small data. Because in our clinic, the spinal cord is contoured not only until L1, L2 how the software does it. But if we are just far below in the area, we just use spinal cord and call the [inaudible 00:03:18.170] as one structure so we have here a problem with the surface points. So I would say there's a...it's complicated to just use this quantitative measurement because for a clinical routine, the structures could have been good as well.
That's why we added a kind of qualitative review where we give every structure with automatically generated score about 1, 2, or 3. And 1 means it's useful without corrections, 2 means it's useful with minor corrections, and 3 means the structure is not useful. Here, have it for the heart. And it looks like the score of 1 correlates with High Dices and the score of 2 even has some High Dices but especially, we have still, maybe the dice of around 0.4 and we can say we can accept it in clinical routine. We've done this for all the structures and on this slide, I just wanna point out that even if we have some problems with, like, the lungs, it was not useful but the dice was still high.
So just to make a decision how good this software is, you have to look at all the quantitative and the qualitative measurements. And I have still for spinal couture here and esophagus. And I would just say in the end that for most of the structures, this software is ready and useful in clinical routine, except for the esophagus where it is not working. And actually, I can add, the version 1.5 is a bit better. Well, I'm still looking forward when we can alternate this, yeah, segmentation process in our clinical routine. Thank you.