Learning from normality

The new approach is based on anomaly detection: From the very precise characterization of normal tissue and findings from frequent diseases, the model learns to recognize and flag deviations, without having to be specifically trained for these rarer cases. For their study, the researchers collected two large datasets of microscopic images of tissue sections from gastrointestinal biopsies with the corresponding diagnoses. In these datasets, the ten most common findings – including normal findings and common diseases such as chronic gastritis – account for around 90 percent of cases, whereas the remaining 10 percent contained 56 disease entities – including many cancers.

For the training and evaluation of their model, the researchers used a total of 17 million histological images from 5,423 cases. “We compared various technical approaches and our best model detected with a high degree of reliability a broad range of rarer pathologies of the stomach and colon, including rare primary or metastasizing cancers. To our knowledge, no other published AI tool is capable of doing this,” says Müller. Using heatmaps, moreover, the AI can indicate in color the position of anomalies in the tissue section.

Significantly easing the diagnosis workload

By identifying normal findings and frequent diseases and detecting anomalies, the new AI model, which will be further improved over time, could provide critical support to doctors. Although the identified diseases still need to be confirmed by pathologists, “doctors can save a lot of time, because normal findings and a certain proportion of the diseases can be automatically diagnosed by AI. This applies to around a quarter to a third of cases,” says Klauschen. “And in the remaining cases, AI can facilitate case prioritization and reduce missed diagnoses. This would represent huge progress.”