ALGORITHM COMBINING VIRTUAL CHROMOENDOSCOPY FEATURES FOR POLYP CLASSIFICATION

Aims Colonoscopy is considered the gold standard for decreasing colorectal cancer incidence and mortality. Visual differentiation between benign and pre-malignant colorectal polyps (CRPs) is an ongoing challenge in clinical endoscopy with accuracies of 71-90% in the Dutch bowel cancer screening program, exposing patients to risks of incorrect optical diagnosis. The PIVI thresholds for CRP-classification are only met in highly selective expert endoscopists. To allow a ‘diagnose-and-leave’ and ‘resect-and-discard’ strategy, optical diagnosis has to be improved. Computer-Aided-Diagnosis systems have been developed to analyze endoscopic images and classifies CRPs by exploiting machine learning algorithms.

The aim of this study is determining the diagnostic accuracy of a novel approach and algorithm for polyp malignancy classification, using one-shot learning with a Triplet Network architecture trained with triplet loss.

Methods The algorithm was trained prospectively by using 609 endoscopic images from 203 polyps. For each polyp, three imaging modalities (Triplet Network architecture) were used to improve algorithm prediction: White light (WL), Blue Laser Imaging (BLI) and Linker Color Imaging (LCI). We performed a retrospective comparative analysis to investigate the accuracy of the algorithm in distinguishing benign polyps (hyperplastic) from pre-malignant lesions (adenomas, sessile serrated lesions), using histopathology as gold standard.

Results 172 polyps were found to be premalignant and 31 were benign polyps. The results of combining the Triplet Network features with additional endoscopic modalities performed an accuracy of 89,7% with a sensitivity of 89,5% and a specificity of 90,3% for polyp malignancy classification. This is comparable to state-of-the-art methods but with much faster inference time (from hours to seconds).

Conclusions Our novel approach and algorithm for automatic polyp malignancy classification differentiates accurately between benign and pre-malign polyps in endoscopic images. This is the first algorithm combining three optical modalities (WL/BLI/LCI) and a Triplet network. The algorithm can be further improved by increasing the amount of images.