Use of Artificial Intelligence Image Classifier to Predict Histology of Solitary Sessile Gastric Lesions and Comparison with Junior Endoscopists

Abstract

Background: Accurate determination of the histology of a sessile gastric lesion usually requires costly multiple biopsies or even technical challenging total enbloc resection. Sampling error from biopsy could also result in false negative result. The use of optical magnified endoscopy in combination with chromoendoscopy or image enhanced endoscopy has been suggested to be able to characterize sessile gastric lesions. Nevertheless, this kind of endoscopic skill is not widely available.

Aim: We evaluated the use of artificial intelligence (AI) assisted image classifier in determining the histology of solitary sessile gastric lesions based on non-magnified narrow band images (NBI).

Methods: An AI image classifier was built on a convolutional neural network with 5 convolutional layers and 3 fully connected layers. It was first trained by 2,000 endoscopic images of both non-neoplastic and neoplastic gastric lesions, as verified by final histology based on either multiple biopsies or total endoscopic resection. The gastric pathology was diagnosed based on WHO classification. Neoplastic lesions were defined as presence of gastric dysplasia, adenoma or carcinoma in the most severe histology of the lesion. Non- neoplastic lesions were defined as the absence of gastric dysplasia, adenoma or carcinoma. The independent validation set then consisted of 170 endoscopic NBI images from 17 gastric lesions. For each gastric lesion, around 10 regions of interest (ROI) were randomly selected in a pixel of at least 300 x 300 dpi (Figure). The endoscopic images were all obtained by non–continuous optical magnifying endoscopy series (GIF-HQ290 model) and CV-290 video system (Olympus Medical system). The independent validation set was also reviewed by three junior endoscopists who had performed more than 2,000 upper endoscopy.

Results: The mean size of the 17 gastric lesions in the validation set was 14.5mm (range: 5 to 25 mm) and 12 (70.6%) were located at the antrum. Six (35.3%) were neoplastic lesions. The performance of AI and the junior endoscopists were summarized in the Table. AI is more confident in the prediction of neoplastic lesions than non- neoplastic lesions (88.7% vs 83.4%, p< 0.01), lesions located at non-antrum area (90.2% vs 83.7 in antrum, p<0.01) and lesions with size > 2cm (92.6% vs 80.7%, p< 0.01). AI was superior to all juniors in term of accuracy 91.2% vs 67.7% or 48.8% or 71.2 (p < 0.001), negative predictive value (NPV) 87.3% vs 56.4% or 40.8% or 60.8% (p < 0.001). and the area under ROC curve (AUROC) 0.93 vs 0.43 or 0.39 or 0.67 (p < 0.001).

Conclusions: The trained AI image classifier based on non- magnified NBI images can accurately predict the presence of neoplastic component within sessile gastric lesions, and its performance is better than junior endoscopists.