Read the Full Video Transcript

Irène Buvat: Good morning, everyone. So Dr. Uribe perfectly set the scene for my presentation. So I'll talk about how artificial intelligence can assist in performing dosimetry. These are my disclosures.

So you are probably well aware that personalized dosimetry requires a complicated path, including several specialty acquisitions, followed by quantitative image reconstruction, registration of images acquired at different time points, delineation of structures of interest.

Then you have to fit the time-activity curve so as to estimate the area under the curve in each voxel or in each region of interest. And then we use a model to deduce the absorbed dose in gray per voxel or per organ.

So this is quite complicated. And this complexity is the current roadblock because as it is considered too complicated by many, dosimetry is rarely performed, so we do not have enough evidence that it is useful. And given that we don't have this evidence, it is not done, and so on. So simplification is really a key to break this catch-22 problem. And artificial intelligence might help.

It can actually help in each and every step, either to increase the accuracy of that step or to make it faster. And it can even allow you to skip some stages, as I will show at the end of the talk. So simplification is absolutely needed for performing more personalized dosimetry studies so that we can fully explore the benefit of that personalized dosimetry.

So let me take a few examples. First, you can use AI to enhance the accuracy of the SPECT images by deblurring these images so that you get rid of some partial volume effect, and you get less biased activity estimates. And because the dose is calculated from the activity estimates, if you improve the activity estimate accuracy, you will improve the accuracy with which you estimate the dose.

So here, for instance, an AI method of partial volume correction. And you can see in the orange curve that you get closer to the true activity profile in blue compared to when you do not perform any partial volume correction.

You can also use deep learning to better perform the registration of the different SPECT-CT images that you acquire at different time points after treatment. And thanks to deep learning, you can actually account for both the SPECT and the CT to perform that registration, instead of only relying on the CT to register the PET scans.

Deep learning is also extremely good at segmentation. And it can be used to automate that segmentation step. This is well illustrated here to segment the bone marrow and estimate the dose to the bone marrow.

So as you can see, the segmentation by the AI is very close to the manual segmentation here. And in this interesting study, they showed a relationship between the dose to the bone marrow here and the platelet counts that was almost identical when the segmentation was performed by an AI or by the expert.

You can also use AI to calculate the dose map directly from the time-integrated activity, without having to go through lengthy Monte Carlo simulation or approximate S values, which are needed to convert an activity map into a dose map. So as you can see here, this is not perfect. But it's pretty accurate even at the voxel level, with errors less than 1 gray basically.

And last, one can use AI to reduce the acquisition time. So if you use proper training, you can infer the neighboring projections from the ones that you acquire. So, for instance, you can acquire only one out of four projections, infer the missing projections, and get a decent reconstruction.

But here, I'd like to give a word of caution because AI is not a magic wand. So you still need counts, and you still need proper sampling of all projections if you want to get accurate activity estimates. You can see that using AI slightly reduced the variability in activity estimates—oops—but it increased the bias. So do not believe these vendors that will tell you that you can get accurate images by reducing the acquisition time by 50, for instance.

And even more tempting, a group from the University of Michigan has trained a model that could estimate the absorbed dose in the tumors only—not in healthy tissues—from a gallium 68 DOTATATE PET-CT without the need to perform SPECT-CT imaging after treatment. So you go directly from the PET pretreatment study up to an estimate of the absorbed dose.

And they found that a simple model based only on the SUV mean in the tumor, SUV mean in the total liver encompassing all lesions, and an SUV mean of all lesions to go where the activity was going could predict pretty well the activity in the tumor. So, of course, you need to have your PET-CT performed just before the treatment—so avoid the gap that Thomas was mentioning yesterday between the pretreatment PET-CT and the treatment.

So to conclude, AI currently appears as a Swiss knife that might greatly facilitate the dosimetry studies in the future to reach that level of evidence we need through prospective trials. Thank you for your attention.