AI-generated images for rare diseases
Menu
Bild: Weltraum und Satellit; Copyright: DLR Bild: Weltraum und Satellit; Copyright: DLR

AI-generated images uncover rare diseases

Significant improvement in diagnostic accuracy

Rare diseases always present medicine with the same problem: a lack of data. In order to detect diseases such as cardiac amyloidosis or somewhat more common bone metastases at an early stage, extensive data sets are needed - but there are hardly any. A research team at MedUni Vienna has found a way to change this: with synthetic image data generated by artificial intelligence.

Clemens Spielvogel und Dr. David Haberl stehen vor einem MRT

Topics in the article:

  • What can the generative AI model do?
  • This is what the synthetic images look like
  • How do you train an AI?
  • Which scintigraphy is real?
  • Quality of the images from a doctor's perspective
  • Results of the study
  • Excursus from an expert's perspective
  • Possibilities and future plans

What can the generative AI model do?

In a study, researchers led by Dr. Clemens Spielvogel (left) and Dr. David Haberl (right) from the Division of Nuclear Medicine (Department of Biomedical Imaging and Image-guided Therapy) at MedUni Vienna have shown that they can artificially recreate typical image patterns of a disease using a generative AI model. The model was trained using image patterns from real patient data. This data can then be used to "feed" algorithms that are used to diagnose the disease. Dr. David Haberl talks about the motivation behind his team's research:

This is what the artificially generated images look like

Six silhouettes - The artificially generated image data can be used to recognize patients with different diseases - copyright: Haberl/Spielvogel
The artificially generated image data show evidence of bone metastases and cardiac amyloidosis.

What is generative AI?

Generative AI is a generic term for systems that can generate new content - texts, images or even medical scans.

On the one hand, generative AI includes language-based models such as ChatGPT. On the other hand, there are image-based models. These include image-to-image translation: one image is transformed into another - such as a CT image from an MRI image.

In addition, fully synthetic medical images can even be generated, as in this case. Spielvogel explains: "In our case, these are scintigraphies and this fully synthetic image data can then be used to create diagnostic models, for example to diagnose diseases at an early stage."

How do you train an AI?

Step by step, the images become more detailed

Scintigraphy scans were used for the training data

Can you recognize which scintigraphy is real?

Bild: real image; Copyright: Spielvogel/Haberl Synthesized image; Copyright: Spielvogel/Haberl

The quality of the images from a docotr's perspective

How realistic the artificially generated image data really is was tested in a blinded survey of doctors. In around 60 percent of cases, the doctors were unable to distinguish whether an image was synthetic or real.

The generated images are really very realistic and true to detail, in the vast majority of cases I couldn't see any difference to real image data. Impressive!

Department of Nuclear Medicine at the Medical University of Vienna
Dr. Dr. Kilian Kluge, Resident Physician

Key findings of the study

As part of the study, an independent research group from the University of Brescia then confirmed the relevance of the synthetic data. They developed the AI system to detect people with suspected cardiac amyloidosis or bone metastases. Afterwards, the system was validated.

Patients from 4 independent institutions served as the basis for the validation
➔ The results showed a significant improvement in diagnostic accuracy through the use of synthetic data.

Excursus from an expert's perspective: the value and risks of AI in diagnostics

MEDICA.de: Dr. Kluge, what advantages do you as a doctor see in this new possibility for diagnosing clinical pictures?

Dr. Kluge: At present, I don't see any direct benefits of AI-generated image data in diagnostic practice. However, since the development of AI applications in medicine is often hampered by a lack of adequate data sets, AI-generated, realistic images could accelerate the development of diagnostic algorithms and thus make innovative applications possible.

MEDICA.de: In your opinion, is there a risk that doctors will rely too heavily on generated data and their own expertise will fade into the background?

Dr. Kluge: This is certainly an important point and a real risk. Training companies and specialist societies should therefore develop concepts aimed at acquiring their own diagnostic skills and the critical use of AI applications.

Possibilities and future plans

The generated images not only help with data scarcity. They can also fill specific gaps, for example in patient groups that are underrepresented in real data sets.

David Haberl adds that the model could also be used to train medical students, for example, as it would be possible to generate specific images of diseases or certain pathologies. He adds: "Synthetic data will never be able to replace real data sets. But I believe it is a promising method, especially when it comes to expanding data sets in a targeted manner."

Smiling woman with long hair - Natascha Mörs; Copyright: beta-web

Author: Natascha Mörs | Editorial team MEDICA-tradefair.com

Multimedia editor Natascha Mörs has been writing and filming for MEDICA.de for almost 20 years. She is particularly interested in personal discussions with experts from the world of medical technology about their practical and increasingly smart solutions.
Bild: Medizinstudentin mit VR-Brille und Controllern; Copyright: UKW/Threedee

Discover more about ...

Read also from Imaging & Diagnostics:

Two medical devices that look like wide bracelets with light-emitting diodes; Copyright: MediSensonic SA

Needle-free diabetes monitoring

Meet the Makers

Head and neck tumor sample taken with a probe; Copyright: Leibniz-IPHT

Cancer surgery: more precise with lasers

Interview

Diagram of a brain that is being scientifically researched

Neurology: holographic imaging

Use Case