Delineation Of Barrett’S Neoplasia Using Deep Neural Networks

Aims Accurate delineation of Barrett’s neoplasia is crucial to detection and planning of endoscopic resection. However, this task can be very challenging. The aim of this study is to develop and validate an artificial intelligence algorithm based on deep neural networks for delineation of Barrett’s neoplasia.

Methods The AI algorithm, based on SegNet architecture, was trained and validated on 75,305 images (94 videos) of neoplastic Barrett’s and 44,586 images (62 videos) of non-neoplastic Barrett’s. For testing, we prospectively recorded videos of histologically confirmed Barrett’s neoplasia independent of the training and validation datasets. Ground truth was histological diagnosis and delineation of neoplasia by 3 experts. AI delineation was assessed against each individual expert’s marking (individual spot), the overlap of all experts marking (optimum spot), as well as the overall area covered by all expert’s markings (total spot).

Tab. 1 Summary of AI performance metrics

	Accuracy	m-precision	m-IoU
Optimum spot	83.3 %	0.43	0.41
Individual spot	100 %	0.62	0.54
Total spot	60.0 %	0.30	0.29

Results We included 30 videos of histologically confirmed Barrett’s neoplasia assessed using all 3 major endoscopy platforms. The AI system correctly detected neoplasia in all videos. We used a 20 % threshold for precision to calculate the accuracy of delineation. The threshold is intended to enable AI-based targeted biopsies. Compared to optimum spot, AI delineated neoplasia lesions with accuracy, mean IoU and mean precision of 83.3 %, 0.41, and 0.43 respectively. When compared to at least one of the individual spots, AI delineated lesions with accuracy, mean IoU and mean precision of 100 %, 0.53 and 0.62 % respectively. [Tab. (1)] summarizes the results.

Conclusions Our deep learning system delineated Barrett’s neoplasia on prospectively recorded endoscopic videos with high accuracy and good overlap compared to expert’s marking. This needs to be validated during real time endoscopy assessment. If proven, this can potentially change the current surveillance protocol from quadratic random biopsies to targeted biopsies and improve the R-0 rates during endoscopic resection.

Citation Abdelrahim M, Saikou M, Maeda N et al. eP483 DELINEATION OF BARRETT’S NEOPLASIA USING DEEP NEURAL NETWORKS. Endoscopy 2021; 53: S256.

Publication History

Article published online:
19 March 2021