Deep Learning Model for Histologic Diagnosis of Dysplastic Barrett's Esophagus: Multisite Cohort External Validation

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Gastrointestinal pathology

Background:

• Barrett's esophagus (BE) carries a risk of progression to esophageal adenocarcinoma (EAC).
• Accurate grading of BE dysplasia is crucial but challenged by interobserver variability in manual pathology reads.
• External validation of a deep learning model for BE dysplasia diagnosis is needed.

Purpose of the Study:

• To externally validate a deep learning model (BEDDLM) for diagnosing dysplasia grades in Barrett's esophagus using whole slide images (WSIs).
• To assess the model's performance across different dysplasia categories: non-dysplastic BE (NDBE), low-grade dysplasia (LGD), and high-grade dysplasia (HGD).

Main Methods:

• Digitization of histology slides from NDBE, LGD, and HGD cases from three external academic centers.
• Normalization of slide stain characteristics using cycle-generative adversarial networks (cGANs).
• Assessment of WSIs by the BEDDLM ensemble model, comprising YOLO and ResNet101 components.

Main Results:

• The study analyzed 489 WSIs with consensus histopathology serving as the criterion standard.
• The BEDDLM ensemble model achieved high sensitivity and specificity across dysplasia grades: NDBE (73.3%, 93.4%), LGD (84.6%, 80.6%), and HGD (80.7%, 94.8%).
• F1 scores indicated strong performance: 0.81 for NDBE, 0.69 for LGD, and 0.83 for HGD.

Conclusions:

• The externally validated deep learning model demonstrates substantial accuracy in diagnosing BE dysplasia grades on whole slide images.
• This AI-driven approach shows promise in improving the consistency and accuracy of BE dysplasia diagnosis.
• The findings support the potential of deep learning in digital pathology for critical gastrointestinal diagnoses.