Recently, I happened to be present at a meeting of professionals in medical imaging where a requirement was formulated concerning the generation of bulk arrays of mock DICOM patients with highly realistic (but faked) data on all levels including the images.

The use cases discussed were the creation of data for bulk tests, especially those of data protection algorithms.

Whereas it is a relatively feasible thing to generate "scalar" entries at all levels down to the series, the generation of images that would look natural and correlate with the data at all levels, but belong to no real patient, seems to be a problem of a different caliber.

I wonder if someone somewhere has tried to apply the technology used in the generation of deepfaked faces to create "deepfaked" DICOM images from the scratch.

So far, I failed to find any paper or mention to this topic. I would be thankful for all information and hints.

Hi,
This is indeed a problem that is more in the realms of medical image processing using deep learning techniques. My colleagues and I have worked a bit on this using generative adversarial networks, such as pix-2-pix or cycle-GANs, for adding fake faces to MR images that may have faces cropped. There are several works already on this that may give you some hints:

https://www.sciencedirect.com/science/article/pii/S1053811921001221
https://www.biorxiv.org/content/10.1101/447102v1.full.pdf
https://arxiv.org/pdf/1810.06455.pdf
https://archive.ismrm.org/2020/0546.html

When you have the images already generated, the creation of the DICOM version of is an easier techincal problem.

All the best with this,

Ricardo
RaC
-- @rcorredorj

On Thursday, March 3, 2022 at 11:14:48 AM UTC+1, francesc...@gmail.com wrote:
> Recently, I happened to be present at a meeting of professionals in medical imaging where a requirement was formulated concerning the generation of bulk arrays of mock DICOM patients with highly realistic (but faked) data on all levels including the images.
>
> The use cases discussed were the creation of data for bulk tests, especially those of data protection algorithms.
>
> Whereas it is a relatively feasible thing to generate "scalar" entries at all levels down to the series, the generation of images that would look natural and correlate with the data at all levels, but belong to no real patient, seems to be a problem of a different caliber.
>
> I wonder if someone somewhere has tried to apply the technology used in the generation of deepfaked faces to create "deepfaked" DICOM images from the scratch.
>
> So far, I failed to find any paper or mention to this topic. I would be thankful for all information and hints.

On Friday, March 4, 2022 at 9:49:16 AM UTC+1, ra.co...@gmail.com wrote:
> Hi,
> This is indeed a problem that is more in the realms of medical image processing using deep learning techniques. My colleagues and I have worked a bit on this using generative adversarial networks, such as pix-2-pix or cycle-GANs, for adding fake faces to MR images that may have faces cropped. There are several works already on this that may give you some hints:
>
> https://www.sciencedirect.com/science/article/pii/S1053811921001221
> https://www.biorxiv.org/content/10.1101/447102v1.full.pdf
> https://arxiv.org/pdf/1810.06455.pdf
> https://archive.ismrm.org/2020/0546.html
>
> When you have the images already generated, the creation of the DICOM version of is an easier techincal problem.
>
> All the best with this,
>
> Ricardo
> RaC
> -- @rcorredorj
> On Thursday, March 3, 2022 at 11:14:48 AM UTC+1, francesc...@gmail.com wrote:
> > Recently, I happened to be present at a meeting of professionals in medical imaging where a requirement was formulated concerning the generation of bulk arrays of mock DICOM patients with highly realistic (but faked) data on all levels including the images.
> >
> > The use cases discussed were the creation of data for bulk tests, especially those of data protection algorithms.
> >
> > Whereas it is a relatively feasible thing to generate "scalar" entries at all levels down to the series, the generation of images that would look natural and correlate with the data at all levels, but belong to no real patient, seems to be a problem of a different caliber.
> >
> > I wonder if someone somewhere has tried to apply the technology used in the generation of deepfaked faces to create "deepfaked" DICOM images from the scratch.
> >
> > So far, I failed to find any paper or mention to this topic. I would be thankful for all information and hints.

Hello Ricardo,
very interesting indeed. I will have a detailed look on that stuff a bit later, but anyway thanks a lot! It is always better to start not from scratch.. I may have a couple a questions after I have read all those papers.
You are right, to transform the generated images into DICOM is a purely technical task.

Best,
FY

On Friday, March 4, 2022 at 2:54:19 PM UTC+1, Francesca Yffelti wrote:
> On Friday, March 4, 2022 at 9:49:16 AM UTC+1, ra.co...@gmail.com wrote:
> > Hi,
> > This is indeed a problem that is more in the realms of medical image processing using deep learning techniques. My colleagues and I have worked a bit on this using generative adversarial networks, such as pix-2-pix or cycle-GANs, for adding fake faces to MR images that may have faces cropped. There are several works already on this that may give you some hints:
> >
> > https://www.sciencedirect.com/science/article/pii/S1053811921001221
> > https://www.biorxiv.org/content/10.1101/447102v1.full.pdf
> > https://arxiv.org/pdf/1810.06455.pdf
> > https://archive.ismrm.org/2020/0546.html
> >
> > When you have the images already generated, the creation of the DICOM version of is an easier techincal problem.
> >
> > All the best with this,
> >
> > Ricardo
> > RaC
> > -- @rcorredorj
> > On Thursday, March 3, 2022 at 11:14:48 AM UTC+1, francesc...@gmail.com wrote:
> > > Recently, I happened to be present at a meeting of professionals in medical imaging where a requirement was formulated concerning the generation of bulk arrays of mock DICOM patients with highly realistic (but faked) data on all levels including the images.
> > >
> > > The use cases discussed were the creation of data for bulk tests, especially those of data protection algorithms.
> > >
> > > Whereas it is a relatively feasible thing to generate "scalar" entries at all levels down to the series, the generation of images that would look natural and correlate with the data at all levels, but belong to no real patient, seems to be a problem of a different caliber.
> > >
> > > I wonder if someone somewhere has tried to apply the technology used in the generation of deepfaked faces to create "deepfaked" DICOM images from the scratch.
> > >
> > > So far, I failed to find any paper or mention to this topic. I would be thankful for all information and hints.
> Hello Ricardo,
> very interesting indeed. I will have a detailed look on that stuff a bit later, but anyway thanks a lot! It is always better to start not from scratch. I may have a couple a questions after I have read all those papers.
> You are right, to transform the generated images into DICOM is a purely technical task.
>
> Best,
> FY

Ricardo, the last paper in the list is only available with membership. Could you please tell me the DOI? Thanks in advance!