AI trained on animal data improves human surgery

Researchers in Germany used medical images from animals to teach an AI model to improve human imaging during surgery, contributing to safer, more precise interventions.  

During surgery, it is difficult to visually distinguish healthy from unhealthy tissue. Surgeons use spectral imaging, which acquires information invisible to the human eye, such as blood flow and oxygen levels, combined with AI analysis, to support their evaluation. However, AI training using human tissue is limited because few images are available, and classical AI models trained on animal data often fail in humans.  

Scientists at the German Cancer Research Center (DKFZ), Heidelberg University Hospital and University Medical Center Mannheim analysed over 14,000 spectral images of humans, pigs and rats, and trained an AI model to identify unhealthy tissue in several organs. A subset of the images was also fully reviewed by two different medical experts and the remaining were analysed by a medical team using an open-source software. 

Instead of focusing on colour patterns alterations in the images, which have been used in classical AI models but differ between animals and humans, the team found common patterns related to oxygen levels and response to blood flow and trained the AI to identify them.  

The AI model could transfer the knowledge across species, since they observed the same patterns in healthy and unhealthy tissue in images from animals and humans. 

“This is an important step toward making surgical procedures safer and more precise in the future,” said co-author Alexander Studier-Fischer, from DKFZ, Heidelberg University Hospital and  University Medical Center Mannheim, who led the clinical aspects of the study published in Nature Biomedical Engineering.  

The AI code and trained models are now available for other researchers to accelerate the introduction of this analysis in clinical settings, where it could be applied even in tissues where no human data is available.