Deep-Learning System Detects Neoplasia in Patients With Barrett's Esophagus With Higher Accuracy Than Endoscopists in...

Area of Science:

• Artificial Intelligence in Medicine
• Gastroenterology
• Oncology

Background:

• Barrett's esophagus (BE) poses a risk for esophageal adenocarcinoma.
• Early detection of neoplasia in BE is crucial for effective treatment.
• Endoscopic detection of early-stage neoplasia in BE remains challenging.

Purpose of the Study:

• To develop and validate a real-time, deep-learning computer-aided detection (CAD) system.
• To enhance the endoscopic detection of early neoplasia in patients with BE.
• To benchmark the CAD system's performance against experienced endoscopists.

Main Methods:

• A hybrid ResNet-UNet deep-learning model was developed for CAD.
• The system was pretrained on 494,364 labeled endoscopic images.
• Validation involved 1704 esophageal high-resolution images from 669 BE patients, with performance compared to 53 endoscopists.

Main Results:

• The CAD system achieved 89% accuracy, 90% sensitivity, and 88% specificity in detecting neoplasia (Dataset 4).
• In Dataset 5, the CAD system (88% accuracy, 93% sensitivity, 83% specificity) outperformed general endoscopists (73% accuracy, 72% sensitivity, 74% specificity).
• CAD demonstrated high accuracy, comparable delineation performance to experts, and identified optimal biopsy sites in over 90% of detected neoplasia.

Conclusions:

• A validated deep-learning CAD system effectively detects neoplasia in BE patients.
• The system offers high accuracy and near-perfect delineation, aiding in early diagnosis.
• This technology represents a significant advancement for primary neoplasia detection in BE.