Researchers from LMU, TU Berlin and Charité have developed a new AI tool that uses imaging data to also detect less common diseases of the gastrointestinal tract.
Artificial intelligence, already used in many areas of medicine, has enormous potential when it comes to helping doctors diagnose diseases with the help of imaging data. However, AI models need to be trained with a large number of examples, which are generally available in sufficient quantities only for common diseases.
As if a family doctor only had to diagnose cough, runny nose and sore throat. The real challenge is also to identify the less common diseases, which current AI models often miss or misclassify.”
Professor Frederick Klauschen, Director of the Institute of Pathology at LMU
Together with the team of Prof. Klaus-Robert Müller from TU Berlin/BIFOLD and colleagues from Charité – Universitätsmedizin Berlin, Klauschen has developed a new approach that overcomes this limitation: As the scientists report in the journal New England Journal of Medicine AI (NEJM AI), their new model only needs training data from common findings to also reliably detect less common diseases. This could significantly improve diagnostic accuracy and reduce the workload of pathologists in the future.
Learning from normality
The new approach is based on abnormality detection: From very precise characterization of normal tissue and findings from common diseases, the model learns to recognize and highlight abnormalities, without having to be specifically trained for these rarer cases. For their study, the researchers collected two large data sets of microscopic images of tissue sections from gastrointestinal biopsies with the corresponding diagnoses. In these data sets, the ten most common findings—including normal findings and common diseases such as chronic gastritis—accounted for about 90 percent of the cases, while the remaining 10 percent contained 56 diseases—including several cancers.
To train and evaluate their model, the researchers used a total of 17 million histological images from 5,423 cases. “We compared several technical approaches and our best model identified with a high degree of reliability a wide range of rarer stomach and colon pathologies, including rare primary or metastatic cancers. To our knowledge, no other published AI tool is able to do this that,” says Müller. Using heat maps, moreover, the AI can indicate the location of abnormalities in the tissue section.
Significantly reducing the diagnosis workload
By identifying normal findings and common diseases and detecting abnormalities, the new AI model, which will be further refined over time, could provide critical support to doctors. Although the detected diseases still need to be confirmed by pathologists, “physicians can save a lot of time because the normal findings and a certain percentage of the diseases can be automatically diagnosed by AI. This is about a quarter to a third of cases,” says Klauschen. “And in the remaining cases, AI can make it easier to prioritize cases and reduce missed diagnoses. That would represent a huge advance.”