Enabling large-scale screening of Barrett's esophagus using weakly supervised deep learning in histopathology

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Barrett's esophagus is a pre-malignant condition for esophageal adenocarcinoma.
• Current diagnostic methods like the Cytosponge-TFF3 test are resource-intensive.
• Large-scale screening is limited by the need for pathologist assessment of multiple slides and biomarkers.

Purpose of the Study:

• To develop a deep learning approach for detecting Barrett's esophagus using only H&E stained slides.
• To eliminate the need for expensive, localized expert annotations in the diagnostic workflow.
• To improve the capacity for large-scale screening of at-risk populations.

Main Methods:

• A deep learning model was trained using diagnostic labels from clinical trial data.
• The model was trained and validated on two independent datasets totaling 1866 patients.
• The approach utilizes routinely stained H&E slides, bypassing the need for TFF3 immunohistochemistry.

Main Results:

• The H&E deep learning model achieved an AUROC of 91.4% on the discovery dataset and 87.3% on the external test dataset.
• Performance was comparable to the TFF3 biomarker model.
• The semi-automated workflow reduced pathologist workload by 48% without compromising diagnostic accuracy.

Conclusions:

• Deep learning on H&E slides offers a viable, less resource-intensive method for Barrett's esophagus detection.
• This AI-driven approach can significantly enhance screening capacity and assist pathologists.
• The findings support the integration of AI into clinical workflows for improved patient outcomes in esophageal cancer prevention.