Synthetic colonoscopy images: AI deceives experienced endoscopists

An international research team from Würzburg and Berlin has developed an AI system that can generate realistic images of colon polyps. In an international study, even experienced endoscopists were often unable to reliably distinguish the AI-generated images from real colonoscopy images. The work, published in “Endoscopy International Open,” examines how realistic synthetic medical image data has become and what potential it holds for training and AI development.

Colonoscopy is central to colon cancer prevention because it allows polyps to be detected and classified. Large amounts of high-quality image data are needed for the training of doctors and for the development of AI-supported assistance systems. The problem is that medical image data is very sensitive, and its dissemination is complex in terms of data protection and ethics. The research team led by Philipp Sodmann and Alexander Hann from the University Hospital Würzburg therefore wanted to check whether synthetic images could be a practical alternative. A study on AI-generated tissue images also reached similar conclusions, for example.

The basis of the work was more than 40 million individual images from over 7,000 colonoscopy examinations from eight centers. With this data, the researchers trained a so-called latent diffusion model, an image AI that can generate high-resolution representations of colon polyps. Images of typical polyp shapes from several Paris classes were generated. However, according to the authors, rare or particularly complex findings cannot yet be reliably generated.

Who can identify AI-generated polyps?

To evaluate the realism of the images, the researchers conducted a blind study. 53 endoscopists from 46 centers in 14 countries participated. The participants were shown 40 images in random order, including 20 real and 20 AI-generated polyp images. For each image, they had to indicate whether it was a real or artificially generated recording and also estimate whether they felt more confident or uncertain about their decision. Interested parties can playfully test whether the polyps shown there are real or artificial at www.thispolypdoesnotexist.com.

The evaluation shows that the artificial images were surprisingly convincing. AI-generated images were correctly identified as artificial in only 66 percent of cases. Real images were correctly classified as real in 80 percent of cases. The overall accuracy of the classifications was 73 percent. It was also noticeable that participants were more often uncertain with AI-generated images and took longer to make their decision. According to the study, some participating doctors reported that they had often had to guess and could only orient themselves by very fine features such as mucosal reflexes or slight blurring.

According to the authors, this indicates that the generated images convincingly replicate essential visual properties of real polyps, such as shape, surface, mucosal structure, and vascular patterns. At the same time, the team investigated whether the AI might simply be reproducing training images. To this end, similarities between training data, real study images, and synthetic images were compared using embedding and distance analyses. The synthetic images shown thus differed from their nearest neighbors in the training dataset, making mere reproduction of the original images used for the study examples unlikely. However, according to the authors, it cannot be completely ruled out that such a model could in principle also generate very similar reproductions.

Potential for training material

The researchers see significant potential in synthetic polyp images for medical training. Such data could supplement training platforms, reduce data protection issues, and improve the availability of image material. “Especially in endoscopy, it is almost impossible to create good training material because photographed precancerous lesions always look different. Here I see great potential for synthetic images to specifically simulate essential properties of precancerous lesions such as size or shape, thereby achieving the greatest learning success for the examiners,” Hann explains to heise online.

(mack)