Missing Metrics for Digitalised Objects-Based Radiotherapy QA: Example of CT-Reference and Captured Surfaces

Mathieu Gonod (00:05):

So my name is Matto. I’m a medical physicist at Centre G.F. Leclerc, France. And I will speak about the missing metrics that can be missing metrics on digitalized objects radiotherapy QA, and how to improve the quality assurance of the surfaces. So the start of our story with Vision RT is based, it starts in 2020 with the first installation of the AlignRT system on the TrueBeam, a second one in 2022 on a TrueBeam. As we have to change our CT, we installed the CT on the new ones in 2023. We have the first installation on our Halcyon. We’ve had the Halcyon in 2023 with the MapRT system. And in 2024, all machines have AlignRT available for patients. And with all of this, we can evaluate and monitor the patient.

Mathieu Gonod (01:15):

We can measure the respiratory cycle of the patient and avoid collision. And all these tools are based on surface acquisition or with pods inside the bunk. And as you can see, AlignRT needs an external body contour to be used as a reference control to install the patient and the MapRT in the RT plan to know where the machine is during the treatment. As I’m a good medical physicist, I want to have a view of risk. And so I want to evaluate the best systems. And so has sometimes very mistakes in the generation. You can, we can start with the external body control generation that is based on different parameters that is that are linked to the TPS and the, and the choice of the services. And sometimes you can have any differences in the volume of this reference surface. And so for AlignRT system, you will have a different functioning for surface acquisition. You can have any problem, for example, surface reconstruction on the first object. And on the bottom, you can have any noise of the surfaces, and the risks for all of this are if you do not perform any IGRT.

Mathieu Gonod (02:59):

To go for those delivery errors because you will have a mispositioning of the patient. If for SimRT you can have 40 construction error and mislocalization of CTV. And from MapRT, if you have a problem with acquisition, you will have collisions. So we need to have a hard quality assurance of these tools. And there’s nothing before 22 that was adapted to the surface acquisition. And most of control that are explained are linked to accuracy or local localization of the phantom. And the thresholds are based on two millimeters. If you are treating for standard radiotherapy and one millimeters, if you are using stereotactic radiotherapy, but nothing about the surface quality checks for digitalized QA objects that are surface acquired by MapRT or about the CT-based external control.

Mathieu Gonod (04:11):

So we need to go for a more advanced QA between the reality and the 3D models. And if you read a little part of the happy and paper, you have information about that for an emerging application, you have to make imperatively fidelity of the acquired surface by the systems. So you need to measure the representative of the system of the acquired surface of reality. And this is name fidelity in the horizon norms and the horizon norm. Describe that you have to preserve the differences between your surfaces and a 3D numerical object reference 3D numerical object before going for metrics. We just have to understand how it’s working in the constructed surface. So you have many points, and all these points are linked by segments and construct any triangle.

Mathieu Gonod (05:22):

And all these triangles will construct a surface. You have to measure the distance between two objects, the surf, the required surface, and the reference object in orange. And you have mainly two different ways to do that. You have to measure the distances between the points, or we measure the difference between the point of the reference surface and the points and the reference surface. You have an example of how software here in the middle that presents the reference surface and the hint map is linked to the distance between the reference surface and the object here. Now that we have introduced how to measure the distance, we can introduce the similarity at six millimeters. That is the percentage of points that are within the six millimeters between the two objects.

Mathieu Gonod (06:34):

And the second one is a kind of measurement of a noise that is called run roughness. And these two items will be used in the next of study. To calculate all of this quantitative matrix, we developed a Python based tool that is able to open ho files and DCM files. We can make the conversion between DCM and Ho, and we have manual and auto registration. And we can, we are able to calculate qualitative and quantitative metrics. We have a registration vector, the similarity and the roughness on all the surfaces or on the, on a part of the surface by a region of interest selection.

Mathieu Gonod (07:27):

The CT was based on three different phantoms. The first one is a physical phantom. It’s the SRS cube from Vision RT with yogurt with water used as a key for registration. We used a second phantom, that is the 3D representation of this first phantom, and for analysis, we used the female torso from Vision RT. For the first 2D, we want to know how it’s the, we want to know the capabilities to make the imaging. So onto our four machines we put the phantom one at the isocenter. We made 10 acquisitions, and we compare this to the phantom two, and we make a SimRT evaluation. The second study is about the roughness of this MapRT acquisition. MapRT have only two pods on the, on the setting, left and right.

Mathieu Gonod (08:34):

So we put the phantom at the centre of the CT. And we move it through the longitudinal field of view of the CT. And we made lateralized analysis. Considering that all the points from the left of the front are linked to the left camera, and from the right of analysis are linked to the right of the camera. The first study is linked to look for the best combination of parameters for the external contours. We position the phantom one on the CT. We make an image, and we generate many external contour based on two parameters, the field threshold and the smoothing. We register. This acquisition is controlled by phantom two, and we made a similarity evaluation for the results.

Mathieu Gonod (09:42):

The study one shows that we have so for each machine, any result we can see that for horizon one, gen five, HD and horizon three, we have a kind of the same results. 79% to 85% of points are less than one millimeter from the reference subject, but one system have only 63% of points. And with this value for those two, two about roughness of acquisition on MapRT, we can see so at the start of the position, so there’s no laser, right? So you can see at the first position of the graph on the first two bars. I don’t know if, okay on the first position, you have a kind of same result between the left side and right side blue and green and more you go inside the c to the foot of the patient.

Mathieu Gonod (10:49):

And do you have any differences in the motion, in the roughness analysis? As you can see in the image on the top right, you have on the left of the image, you have normal noise on the surface, but on the right you have a very noisy surface. So we asked Vision RT, they came and they found that there is any moisture on the optical field filter. So they replace, and we have no more of this problem about the best combination of parameters for the generation of external contour. You have a graph with similarity on the left at one millimeters, so the percentage of points that are less than one millimeters in function of the smoothing. You have the representation for three different thresholds. As you can see on the top left, you have the best optimal parameters that have, which are given for a minus five 30 field threshold, and we are missing one with a similarity at one millimeters of 90% of a point to compare to our actual generation parameter, with a similarity at 78% and 39%.

Mathieu Gonod (12:13):

We designed this study with this phantom, but we want to go further by using a more representative phantom than a patient to confirm these results. So we developed a Python-based QA tool that is able to manipulate OB today fast for visualization conversion registration. We are able to calculate qualitative and quantitative comparison as similarity and robustness metrics that are reproducible and quantitative. With this, we are able to detect suboptimal surface fast quality, and we can monitor them to trigger performance check calibration, preventive maintenance, or anything else about the CT base contour. We can, we are able to optimize our parameters, but we just want to confirm on a different phantom. And we understand that for digitalized object as your faces, and especially if you have any predictive tools as MapRT, you have a quality assurance that is critical for the subjects.

Mathieu Gonod (13:32):

And we think that similarity and roughness are more necessary and complementary than the actual accuracy tests that are published. So, about the software, we want to include other manufacturers’ files, and we want in 26 to implement all these QA metrics in the QA program. We want to have the system inside, and we want to benchmark to have an installation on fast and easy acquisition. And we have many clinical perspectives. As you can see with the software, it’s QA, but you can manipulate the DICOM object. So you are able to, for example, evaluate the immobilization device deformation or settling. You can correct for MapRT TPS patient position. As you can see here in the image at the bottom, the second image is the DICOM structures of the patient’s body with the patient harms help, but the MapRT acquisition was just a few seconds after the patient put his arm down.

Mathieu Gonod (14:49):

So the collision evidence map is not predictive, and on the software, you have the possibilities to superimpose DICOM to surfaces. You can calculate here the differences between the patient and these two objects. And as you can see in the middle image, you have many reds on the shoulders of the patient. So it means that you have more one centimeters here from the two positions. So you have a problem in the positioning of the patient. And another thing that can, we can do is insert the surface files into head clips to calculate any dose because there is any modification of the patient based on the loss, any volume, I don’t haven’t the word, sorry. And you have many other solutions based on this to, so this kind of software will lead to many other clinical perspectives to improve the quality and safety of the treatments.