Researchers have developed a new artificial intelligence (AI) system that helps doctors more accurately grade Barrett’s esophagus, a condition that can lead to esophageal cancer. This advancement is important because current methods for assessing the severity of abnormal cell changes in Barrett’s esophagus—called dysplasia—often vary between pathologists, especially when identifying early, low-grade changes. The new AI tool aims to reduce this uncertainty and support better clinical decisions without needing complicated image preprocessing steps that have limited previous AI approaches.
Key Takeaways
- The AI system analyzed nearly 1,000 pathology slides from four different medical centers, showing strong accuracy in classifying Barrett’s esophagus tissue as non-dysplastic, low-grade, or high-grade dysplasia.
- It achieved a high overall accuracy score (AUC of 0.96) on a separate test set, meaning it can reliably distinguish between different grades of dysplasia.
- The AI demonstrated excellent sensitivity (98%) for detecting dysplasia, correctly identifying almost all positive cases, while maintaining good specificity (81%) for ruling out non-dysplastic samples.
- Unlike prior systems, this model worked well across multiple institutions without requiring stain normalization—a common but complex image adjustment step—making it more adaptable to real-world clinical settings.
The researchers tackled the challenge of variability in dysplasia grading by training an AI system on a large, diverse set of digitized tissue samples stained with a standard technique called hematoxylin and eosin (H&E). Instead of relying on traditional image preprocessing to standardize the appearance of slides from different labs, the team used a “foundation model” approach. Foundation models are large, pre-trained AI systems that can understand complex patterns across varied data. Here, the model learned to recognize subtle tissue features directly, even with differences in staining between institutions.
The AI workflow involved two main steps. First, “patch-level classifiers” scanned small regions of each slide to identify areas that looked diagnostically important. Then, a technique called “multiple instance learning” combined information from these patches to make an overall diagnosis for the entire slide. The system was also “attribute-aware,” meaning it took into account specific characteristics of the tissue to improve its slide-level predictions. This approach allowed the AI to mimic how expert pathologists focus on key regions when making their assessments.
Testing on a separate set of 130 slides showed the AI’s performance was consistent and robust. Notably, the model never misclassified high-grade dysplasia as non-dysplastic Barrett’s esophagus, which is critical since missing severe cases could delay necessary treatment. The high sensitivity and strong agreement with expert consensus suggest the system could serve as a valuable second opinion or screening tool to reduce diagnostic variability.
By eliminating the need for stain normalization, this AI tool is potentially easier to implement across different hospitals and pathology labs, which often use varying staining protocols. This flexibility could accelerate adoption and improve the consistency of Barrett’s esophagus diagnosis worldwide.
Looking ahead, the researchers envision further testing in prospective clinical trials to evaluate how the AI performs in routine practice. If successful, this technology could help standardize dysplasia grading, guide treatment decisions more reliably, and ultimately improve patient outcomes by catching precancerous changes earlier and more accurately.
Based on research published on arXiv by Amirali Khosravi, Jaidip M. Jagtap, D. Chamil Codipilly et al..
