Researchers have introduced an innovative AI tool that leverages imaging data to identify less commonly occurring diseases related to the gastrointestinal system.
A collaborative effort by researchers from LMU, TU Berlin, and Charité has resulted in the creation of a new AI tool designed to detect rare gastrointestinal diseases using imaging data.
Artificial intelligence is already making waves in various medical fields, showcasing great promise in aiding healthcare professionals with disease diagnoses through imaging data. However, developing AI models requires a substantial number of training examples, which are primarily available for more prevalent diseases. “It’s similar to a primary care doctor only having to identify symptoms like coughs, colds, and sore throats,” states Professor Frederick Klauschen, Director at the LMU Institute of Pathology. “The real challenge lies in identifying those less common ailments, which current AI models frequently miss or misidentify.”
To tackle this issue, Klauschen, in collaboration with Professor Klaus-Robert Müller from TU Berlin/BIFOLD and colleagues at Charité — Universitätsmedizin Berlin, has created a groundbreaking method that addresses this gap: According to their findings published in the journal New England Journal of Medicine AI (NEJM AI), their new model can accurately detect rare diseases with training based solely on more common cases. This advancement could greatly enhance diagnostic precision and alleviate the burden on pathologists in the future.
Learning from Norms
The innovative method centers around anomaly detection. By thoroughly defining normal tissue characteristics and documenting common disease presentations, the model learns to identify and highlight irregularities, without needing explicit training for those rarer instances. For their research, the team gathered two extensive datasets of microscopic images from gastrointestinal biopsy tissue samples, which came with their respective diagnoses. In these datasets, the ten most frequent findings—covering both normal tissue and common diseases like chronic gastritis—constitute roughly 90 percent of the cases, with the remaining 10 percent containing 56 different disease types, including various cancers.
For their model’s training and assessment, the researchers utilized a total of 17 million histological images sourced from 5,423 cases. “We analyzed different technical strategies, and our most effective model reliably identified a wide array of rarer stomach and colon pathologies, including uncommon primary tumors and metastatic cancers. As far as we know, no other AI tool published previously has been able to achieve this,” Müller explains. Additionally, the AI employs heatmaps to visually represent the location of anomalies within the tissue section.
Significantly Reducing Diagnostic Workload
By recognizing normal findings and frequent diseases while also detecting anomalies, this new AI model, expected to improve further with time, could offer essential assistance to healthcare providers. Although the pathologist must still verify the identified diseases, “doctors can greatly reduce their time spent, as the AI can automatically diagnose normal readings and some disease cases. This is applicable to about a quarter to a third of instances,” Klauschen notes. “In cases that remain, AI can aid in prioritizing them and minimizing missed diagnoses, signifying a significant advancement.”
