Video-based computer aided detection system improves Barrett’s neoplasia detection of general endoscopists in a multi-step benchmarking study

Aims Timely endoscopic detection of Barret's neoplasia has significant influence on patient outcome. Computer Aided Detection (CAD) systems may assist in neoplasia detection.

Methods The system was pretrained with ImageNet followed by domain-specific pretraining with GastroNet. GastroNet comprises >5 million endoscopic images. The system was then trained and validated on a BE dataset originating from 15 international endoscopy centers, comprising 6.337 neoplastic (1.362) patients) and 7.695 non-dysplastic images (1.139 patients). All images had histopathological confirmation. Neoplastic images were delineated by expert endoscopists. The system was tested on two prospective video test sets. The test set 1, comprising 71 neoplastic (45 cases) and 180 non-dysplastic (66 cases) videos, included all consecutive cases acquired from January to March 2022. Test set 2 comprised 47 neoplastic (47 cases) and 141 non-dysplastic (82 cases) videos and was enriched with subtle cases of neoplasia. Test set 2 was evaluated by 63 general endoscopists in without and with ADe assistance. Finally, 14 international, independent BE experts evaluated this test set.

Results Sensitivity and specificity of the CAD system were 97% and 85% for test set 1 and 91% and 82% for test set 2. Sensitivity of general endoscopists increased from 67% to 79% with CAD assistance, whilst specificity decreased from 96% to 94%. The sensitivity and specificity of experts were 86% and 90%, respectively ([Table 1]).

Table 1  

Conclusions CAD outperformed general endoscopists in detecting BE neoplasia. Providing CAD to general endoscopists significantly improves their detection rate. CAD detects virtually all neoplasia in a test set representing daily practice. CAD has a detection rate on par with experts.

Table 1) Performance of general endoscopists, expert endoscopists and the CAD system on test set 2

Conflicts of interest

This research has received logistical and financial support from Olympus Tokyo.

Publication History

Article published online:
14 April 2023