Detecting early neoplasia in Barrett’s oesophagus is challenging due to its subtle endoscopic appearance. The Barrett’s Oesophagus Imaging for Artificial Intelligence (BONS-AI) consortium developed a Computer-Aided Detection (CADe) system to aid in identifying neoplastic lesions efficiently.
The study, published in The Lancet Digital Health, integrates prospectively collected data and compares CADe performance with that of Barrett’s oesophagus experts and general endoscopists.
Methodology Overview
Optimization and Threshold Selection: The CADe system’s parameters and hyperparameters were fine-tuned using a prospective dataset of 200 Barrett’s oesophagus images.
Performance Evaluation: Two independent test sets were created: an all-comers set reflecting daily clinical practice and a benchmarking set enriched with challenging cases.
Comparison with Experts: Barrett’s oesophagus experts evaluated the benchmarking set to establish a performance reference.
Impact on Endoscopists: General endoscopists assessed the benchmarking set in two phases: without and with CADe assistance.
Key Findings
CADe Threshold: The threshold for neoplasia detection was set at 0.35.
Performance: CADe showed high sensitivity (95% for images, 97% for videos) and moderate specificity (70% for images, 85% for videos) in the all-comers set.
Comparison with Endoscopists: CADe outperformed general endoscopists in sensitivity but had lower specificity.
Impact on Endoscopists: CADe significantly increased sensitivity while maintaining stable specificity in both image and video test sets.
Lesion Localization: CADe achieved 100% correct localization of neoplastic lesions in images.
Conclusion
The CADe system offers valuable assistance to general endoscopists in detecting neoplastic lesions in Barrett’s oesophagus. Further optimization is necessary for specificity improvement and lesion localization. Nonetheless, CADe integration holds promise for enhancing early neoplasia detection and management.
