Evaluation of an improved computer-aided detection system for Barrett's neoplasia in real-world imaging conditions

Computer-aided detection (CADe) systems may improve detection of Barrett's neoplasia. Most CADe systems are developed with data from expert centers, unrepresentative of heterogeneous imaging conditions in community hospitals, and therefore may underperform in routine practice. We aimed to develop a robust CADe system (CADe 2.0) and compare its performance to a previously published system (CADe 1.0) under heterogeneous imaging conditions representative of real-world clinical practice.

CADe 2.0 was improved through a larger and more diverse training dataset, optimized pretraining, data augmentation, ground truth use, and architectural adjustments. CADe systems were evaluated using three prospective test sets. Test set 1 comprised 428 Barrett's videos (114 patients across five referral centers). Test set 2 addressed endoscopist-dependent variation (e. g. mucosal cleaning and esophageal expansion), with paired subsets of high, moderate, and low quality images (122 patients). Test set 3 addressed endoscopist-independent variation, with 16 paired subsets of 396 images (122 patients), each being based on a different software image-enhancement setting.

CADe 2.0 outperformed CADe 1.0 on all three test sets. In test set 1, sensitivity increased significantly from 87 % to 96 % (P = 0.02), while specificity remained comparable (73 % vs. 74 %; P = 0.73). In test set 2, CADe 2.0 consistently surpassed CADe 1.0 across all image quality levels, with the largest performance gains observed on lower quality images (sensitivity 78 % vs. 61 %; specificity 89 % vs. 77 %; area under the curve 89 % vs. 75 %). In test set 3, CADe 2.0 showed improved performance and displayed reduced performance variability across enhancement settings.

Based on several key improvements, CADe 2.0 demonstrated increased detection rates and better robustness to data heterogeneity, making it ready for clinical implementation.