Subscribe to RSS
DOI: 10.1055/s-0040-1704154
COMPARISON OF THE PRECISION OF OPTICAL DIFFERENTATION BETWEEN NEOPLASTIC AND NON-NEOPLASTIC SUBCENTRIMETRIC POLYP HISTOLOGY BY ENDOSCOPIC EXPERTS AND ARTEFICIAL INTELLIGENCE DEEP LEARNING NEURAL NETWORK (POLYPBRAIN)
Publication History
Publication Date:
23 April 2020 (online)
Aims Background Precise differentiation between sub-centimetric neoplastic and non-neoplastic polyps is important. New polyp classification systems such as BASIC classification and artificial intelligence deep learning algorithm-based Decision Support System (AI-DSS) are developed to achieve high accuracy necessitated by application of the resect-and-discard strategy in everyday practice for non-expert endoscopists.
Our aim was to compare the performance and accuracy of our recently developed AI-DSS to five colonoscopic experts (more than 2000 colonoscopy per year and 20000 during their carrier) familiar with BASIC classification in the differentiation between neoplastic (adenomas) and non-neoplastic (hyperplastic lesions) sub-centimetric polyps.
Methods AI-DSS was trained on our anonymous electronic database from a total of 1800 histologically identified sub-centimetric colorectal polyps and 26000 HD, electronic chromo-endoscopic images. All polyps were removed and sent for histology for final diagnosis. We excluded malignant, juvenile, inflammatory and sessile serrated polyps from the current study protocol. Test set contained 61 HQ pictures from randomly selected polyps (31 neoplastic and 30 non-neoplastic) that was made with Blue Light Imaging (BLI) combined with 50 times optical zoom technology. We made sure that the same polyp’s images were not selected to both train and test sets.
Results In the prediction of histology, AI-DSS versus experts achieved similarly excellent results without significant differences in accuracy, sensitivity, specificity, PPV and NPV: 96.72% vs. 93.03%, 100% vs. 92.74%, 93.55% vs 93.34%, 100% vs. 93.85% and 93.75% vs. 93.09%, respectively. There was a good inter-observer agreement according to the BASIC classification predictors in each polyp and this was correlated with the final histology well.
Conclusions The potential of using our AI-DSS Polypbrain is that it could provide a highly accurate tool for real-time optical diagnosis of neoplastic and non-neoplastic polyps at a similar precision level as high quality expert endoscopists. (Our study was supported by EU Grant: GINOP 2.1.1.-15-2015-00128).
#