Deep Neural Network for the Localisation of Early Neoplasia in Barrett’s Oesophagus with Targeted Biopsies

Aims To develop a deep neural network to diagnose and localise dysplasia in Barrett’s oesophagus (BE).

Methods Videos were collected in high definition white light/optical chromoendoscopy with i-scan (Pentax Hoya, Japan) imaging modes in patients with dysplastic BE lesions (high grade dysplasia (HGD)/intramucosal adenocarcinoma) and non-dysplastic BE (NDBE). Videos were annotated for presence/absence of dysplasia. These were histologically confirmed. We trained a convolutional neural network with a Resnet101 architecture to classify images using annotated video frames. We trained a second network with a FCNResnet50 architecture with the same case split using expert delineations (published dysplasia detection rate of > 90 %) on high quality images to generate targeted biopsy predictions.

Results 124 patients with a video of BE assessment (68 HGD/Intramucosal cancer and 56 controls (54 NDBE, 2 normal oesophagus)) were included. Cases were divided into three independent training, validation and testing groups. The network was trained using 148,936 frames. This was tested on 6 high quality images per case (168 iscan-1 images from 28 dysplastic patients, 102 iscan-1 images from 17 NDBE patients).

The neural network classified BE dysplasia with a sensitivity of 90.5 %, specificity of 80.4 %, area under the ROC was 93.5 %. Heat maps generated from the classifier algorithm overlapped with at least one expert ground truth delineation in 98 % of the test set images where a true positive diagnosis of dysplasia was made.

A maximum of two targeted biopsies were predicted by the second neural network for all 28 dysplastic patients. 91 % of biopsies were correctly within the union of expert delineation in all images.

Conclusions Our neural networks can classify/localise dysplastic BE with targeted biopsies with high accuracy matching experts. The classifier was created using video annotations minimising selection bias. This can potentially minimise need for Seattle protocol biopsies, save costs and time maximising endoscopy capacity.

Citation: Hussein M, Puyal JG-B, Lines D et al. OP10 DEEP NEURAL NETWORK FOR THE LOCALISATION OF EARLY NEOPLASIA IN BARRETT’S OESOPHAGUS WITH TARGETED BIOPSIES. Endoscopy 2021; 53: S8.

Publication History

Article published online:
19 March 2021